Awesome LLMs Datasets Download - Awesome LLMs Datasets Código fuente de descarga

Impresionante-llms-datasets

Resume conjuntos de datos de texto de LLMS representativos existentes a través de cinco dimensiones: corpus de pre-entrenamiento, conjuntos de datos de instrucciones de ajuste fino, conjuntos de datos de preferencias, conjuntos de datos de evaluación y conjuntos de datos NLP tradicionales . (Actualizaciones regulares)
Se han agregado nuevas secciones de conjuntos de datos: conjuntos de datos de modelos de lenguaje grande (MLLMS) multimodal (MLLMS), conjuntos de datos de generación aumentada de recuperación (RAG) . (Actualizaciones graduales)

Papel

Se ha publicado el documento "conjuntos de datos para modelos de idiomas grandes: una encuesta integral" . （2024/2）

Abstracto:

Este documento se embarca en una exploración en los conjuntos de datos del modelo de lenguaje grande (LLM), que juegan un papel crucial en los avances notables de los LLM. Los conjuntos de datos sirven como infraestructura fundamental análoga a un sistema de raíz que sostiene y fomenta el desarrollo de LLM. En consecuencia, el examen de estos conjuntos de datos surge como un tema crítico en la investigación. Para abordar la falta actual de una visión general integral y un análisis exhaustivo de los conjuntos de datos LLM, y para obtener información sobre su estado actual y tendencias futuras, esta encuesta consolida y clasifica los aspectos fundamentales de los conjuntos de datos LLM desde cinco perspectivas: (1) Pre-PRE- capacitación de corpus; (2) instrucción de conjuntos de datos ajustados; (3) conjuntos de datos de preferencias; (4) conjuntos de datos de evaluación; (5) conjuntos de datos de procesamiento de lenguaje natural tradicional (PNL). La encuesta arroja luz sobre los desafíos predominantes y señala posibles vías para futuras investigaciones. Además, también se proporciona una revisión exhaustiva de los recursos de conjunto de datos disponibles existentes, incluidas estadísticas de 444 conjuntos de datos, que cubren 8 categorías de idiomas y abarcan 32 dominios. La información de 20 dimensiones se incorpora a las estadísticas del conjunto de datos. El tamaño total de datos encuestado supera los 774.5 TB para los corpus de pre-entrenamiento y 700m instancias para otros conjuntos de datos. Nuestro objetivo es presentar todo el panorama de los conjuntos de datos de texto de LLM, que sirve como una referencia integral para los investigadores en este campo y contribuye a futuros estudios.

Fig. 1. La arquitectura general de la encuesta. Zoom para una mejor vista

Módulo de información del conjunto de datos

El siguiente es un resumen del módulo de información del conjunto de datos.

Nombre del cuerpo/conjunto de datos
Editor
Tiempo de lanzamiento
- "X" indica un mes desconocido.
Tamaño
Público o no
- "Todos" indica un código abierto completo;
- "Parcial" indica código parcialmente abierto;
- "No" indica no código abierto.
Licencia
Idioma
- "En" indica inglés;
- "Zh" indica chino;
- "AR" indica árabe;
- "Es" indica español;
- "Ru" indica ruso;
- "De" indica alemán;
- "Ko" indica coreano;
- "LT" indica lituano;
- "FA" indica persa/farsi;
- "PL" indica lenguaje de programación;
- "Multi" indica multilingüe, y el número entre paréntesis indica el número de idiomas incluidos.
Método de construcción
- "HG" indica corpus/conjunto de datos generado por humanos;
- "MC" indica Corpus/DataSet construido con el modelo;
- "CI" indica colección y mejora del corpus/conjunto de datos existente.
Categoría
Fuente
Dominio
Categoría de instrucciones
Método de evaluación de preferencias
- "Vo" indica voto;
- "Entonces" indica clasificar;
- "SC" indica puntuación;
- "-H" indica conducido por humanos;
- "-M" indica que se realizan por modelos.
Tipo de pregunta
- "SQ" indica preguntas subjetivas;
- "OQ" indica preguntas objetivas;
- "Multi" indica múltiples tipos de preguntas.
Método de evaluación
- "CE" indica la evaluación del código;
- "Él" indica evaluación humana;
- "Me" indica la evaluación del modelo.
Enfocar
Número de categorías/subcategorías de evaluación
Categoría de evaluación
Número de categorías de entidades (tarea ner)
Número de categorías de relaciones (RE Tarea)

Colegio de cambios

（2024/01/17） Cree el repositorio de datos de datos Awesome-Llms-Datasets .
（2024/02/02） Revise la información para algunos conjuntos de datos; Agregue DOLMA (corpus de pre-entrenamiento | Corporación general de pre-entrenamiento | Categoría multicatina).
（2024/02/15） Agregue la colección AYA (instrucciones de datos de ajuste fino | Instrucción general Inserción de datos de ajuste | Hg & Ci & MC); Conjunto de datos AYA (Instrucción de conjuntos de datos ajustados | Instrucción general de instrucciones de datos ajustados | Hg).
（2024/02/22） Agregue OpenMathInstruct-1 (instrucciones de datos de ajuste fino | Instrucción específica de dominio conjuntos de datos ajustados | Matemáticas); Finben (conjuntos de datos de evaluación | Financiero).
（2024/04/05）
- Agregue nuevas secciones del conjunto de datos: (1) conjuntos de datos de modelos de lenguaje grande multimodal (MLLMS); (2) conjuntos de datos de generación aumentada de recuperación (RAG) .
- Agregar MMRS-1M (conjuntos de datos MLLMS | Instrucción de conjuntos de datos ajustados); VideoChat2-IT (MLLMS DataSets | Instrucción ajustando los conjuntos de datos); InstructDoc (MLLMS DataSets | Instrucción ajustando los conjuntos de datos); Datos Allava-4V (conjuntos de datos MLLMS | Instrucción ajustando los conjuntos de datos); Mvbench (conjuntos de datos MLLMS | conjuntos de datos de evaluación); Olympiadbench (conjuntos de datos MLLMS | conjuntos de datos de evaluación); MMMU (conjuntos de datos MLLMS | conjuntos de datos de evaluación).
- Agregue la serie de referencia de Clue (conjuntos de datos de evaluación | Plataforma de evaluación); OpenLLM Raeperboard (conjuntos de datos de evaluación | Plataforma de evaluación); OpenCompass (conjuntos de datos de evaluación | Plataforma de evaluación); Tablero de clasificación MTEB (conjuntos de datos de evaluación | Plataforma de evaluación); Tablero de clasificación C-MTEB (conjuntos de datos de evaluación | Plataforma de evaluación).
- Agregar NAH (aguja-in a-haystack) (conjuntos de datos de evaluación | texto largo); Tooleyes (conjuntos de datos de evaluación | herramienta); Uhgeval (conjuntos de datos de evaluación | Factualidad); Clongeval (conjuntos de datos de evaluación | Texto largo).
- Agregar MathPile (corpus de pretrabenamiento | Corporos de pre-entrenamiento específicos del dominio | Matemáticas); Wanjuan-CC (Corporos de pre-entrenamiento | Corporativos de pre-entrenamiento general | Páginas web).
- Agregue IEPILE (instrucciones de datos de ajuste fino | Instrucción general Autor de datos ajustados | CI); Instructie (instrucciones de datos de ajuste fino | Instrucción general Autor de datos ajustados | Hg).
- Agregue Crud-Rag (RAG DataSets); Wikieval (conjuntos de datos de trapo); RGB (RAG DataSets); RAG-INSTRUST-BENCHMING-TESTER (RAG DataSets); Ares (conjuntos de datos de trapo).
（2024/04/06）
- Agregar GPQA (conjuntos de datos de evaluación | Asunto); MGSM (conjuntos de datos de evaluación | multilingüe); Halueval-Wild (conjuntos de datos de evaluación | Factualidad); CMATH (conjuntos de datos de evaluación | Asunto); Finemath (conjuntos de datos de evaluación | Asunto); QA en tiempo real (conjuntos de datos de evaluación | Factualidad); WyWeb (conjuntos de datos de evaluación | Asunto); ChineseFactEval (conjuntos de datos de evaluación | Factualidad); Estrellas de conteo (conjuntos de datos de evaluación | Texto largo).
- Agregar Slimpajama (Corporos de pre-entrenamiento | Corporación general de pre-entrenamiento | Categoría Multi-Categoría); MassivExtext (corpus de pretrabenamiento | Corporación general de pre-entrenamiento | Multi-categoría); Madlad-400 (corpus de pre-entrenamiento | Corporativos de pre-entrenamiento general | Páginas web); Minerva (corpus de pre-entrenamiento | Corporación general de pre-entrenamiento | Multi-categoría); Ccaligned (corpus de pre-entrenamiento | corpus de pre-entrenamiento general | corpus paralelo); Wikimatrix (corpus de pre-entrenamiento | corpus de pretrabantes generales | corpus paralelo); OpenWebMath (corpus previos a la capacitación | Corporos previos al entrenamiento específicos del dominio | Matemáticas).
- Agregue WebQuestions (conjuntos de datos NLP tradicionales | Respuesta de preguntas | Conocimiento QA).
- Agregue Alce (conjuntos de datos RAG).
- Agregar alfafin (instrucciones de datos de ajuste fino | Instrucción específica del dominio conjuntos de datos ajustados | Otro); CoIG-CQIA (Instrucción de conjuntos de datos ajustados | Instrucción general Autor de datos ajustados | HG & CI).
（2024/06/15）
- Agregar pista (conjuntos de datos de evaluación | Médico); Bench CHC (conjuntos de datos de evaluación | General); CIF Bench (conjuntos de datos de evaluación | General); Aclue (conjuntos de datos de evaluación | Asunto); LESC (conjuntos de datos de evaluación | NLU); AlignBench (conjuntos de datos de evaluación | Multitasa); Sciknoweval (conjuntos de datos de evaluación | Asunto).
- Agregar MAP-CC (Corporos de pre-entrenamiento | Corporos de pre-entrenamiento general | Categoría Multi-Categoría); FineWeb (corpus de pre-entrenamiento | Corporativos de pre-entrenamiento general | Páginas web); CCI 2.0 (Corporos de pre-entrenamiento | Páginas web de pre-entrenamiento general | Páginas web).
- Agregue WildChat (instrucción de conjuntos de datos ajustados | MC).
- Agregar OpenHermesPreferences (conjuntos de datos de preferencias | sort); HUOZI_RLHF_DATA (PREFERENCIA DATASETS | VOTO); HelpSteer (PREFERENCIA DE DATASTS | SCUENTA); HelpSteer2 (conjuntos de datos de preferencias | puntaje).
- Agregar Bench MMT (conjuntos de datos MLLMS | conjuntos de datos de evaluación); Moscar (conjuntos de datos MLLMS | Corporativos de pre-entrenamiento); MM-NIAH (conjuntos de datos MLLMS | conjuntos de datos de evaluación).
- Agregue Crag (conjuntos de datos RAG).
（2024/08/29）
- Agregar GameBench (conjuntos de datos de evaluación | razonamiento); Halludial (conjuntos de datos de evaluación | Factualidad); Wildbench (conjuntos de datos de evaluación | General); DomaineVal (conjuntos de datos de evaluación | Código); Sysbench (conjuntos de datos de evaluación | General); Kobest (conjuntos de datos de evaluación | NLU); Sarcasmbench (conjuntos de datos de evaluación | NLU); C ³ Banco (conjuntos de datos de evaluación | Asunto); TableBench (conjuntos de datos de evaluación | razonamiento); ArableGaleval (conjuntos de datos de evaluación | Ley).
- Agregar multiprusto (conjuntos de datos MLLMS | conjuntos de datos de evaluación); Obelisc (conjuntos de datos MLLMS | Corporativos de pre-entrenamiento); Multimed (conjuntos de datos MLLMS | conjuntos de datos de evaluación).
- Agregue DCLM (corpus de pre-entrenamiento | Páginas web de pre-entrenamiento general | Páginas web).
- Agregue lituaniano-qa-v1 (instrucciones de datos de ajuste fino | CI & MC); Reinstrucción (instrucción de conjuntos de datos ajustados | HG & CI & MC); Kollm-Converations (instrucciones de datos de ajuste fino | CI).
（2024/09/04）
- Agregue LongWriter-6k (instrucciones de datos de ajuste fino | CI y MC).
- Agregar Medtrinity-25m (conjuntos de datos MLLMS | conjuntos de datos de evaluación); MMIU (conjuntos de datos MLLMS | conjuntos de datos de evaluación).
- Agregue expository-prosa-v1 (corpus de pretrabenamiento | corporativos de pre-entrenamiento general | categoría multicategoría).
- Agregar debateqa (conjuntos de datos de evaluación | Conocimiento); Aguja de aguja (conjuntos de datos de evaluación | texto largo); Arabicmmlu (conjuntos de datos de evaluación | Asunto); PersianmMlu (conjuntos de datos de evaluación | Asunto); TMMLU+ (conjuntos de datos de evaluación | Asunto).
- Agregar Rageval (RAG DataSets); LFRQA (RAG DataSets); Multihop-Rag (conjuntos de datos RAG).
Lanzaremos la información del conjunto de datos en formato CSV.

Tabla de contenido

Corpus para capacitar
- Corporativos de pre-entrenamiento general
  - Página web
  - Textos del idioma
  - Libros
  - Materiales académicos
  - Código
  - Corpus paralelo
  - Redes sociales
  - Enciclopedia
  - Multicatategoría
- Corpus de pre-entrenamiento específicos del dominio
  - Financiero
  - Médico
  - Matemáticas
  - Otro
Instrucciones de datos de ajuste fino
- Instrucción general para conjuntos de datos ajustados
  - Conjuntos de datos generados por humanos (HG)
  - Conjuntos de datos construidos con modelo (MC)
  - Colección y mejora de los conjuntos de datos existentes (CI)
  - HG y CI
  - HG y MC
  - CI y MC
  - HG & CI y MC
- Instrucción específica de dominio conjuntos de datos ajustados
  - Médico
  - Código
  - Legal
  - Matemáticas
  - Educación
  - Otro
Conjuntos de datos de preferencias
- Métodos de evaluación de preferencias
  - Votar
  - Clasificar
  - Puntaje
  - Otro
Conjuntos de datos de evaluación
- General
- Examen
- Sujeto
- NLU
- Razonamiento
- Conocimiento
- Texto largo
- Herramienta
- Agente
- Código
- Ácido
- Ley
- Médico
- Financiero
- Normas sociales
- Realidad
- Evaluación
- Multitarea
- Plurilingüe
- Otro
- Plataforma de evaluación
Conjuntos de datos tradicionales de PNL
- Respuesta de preguntas
  - Comprensión de lectura
    - Selección y juicio
    - Prueba de cloze
    - Respuesta a la extracción
    - QA sin restricciones
  - Conocimiento QA
  - Razonamiento QA
- Reconocimiento de la implicación textual
- Matemáticas
- Resolución de coreferencia
- Análisis de sentimientos
- Emparejamiento semántico
- Generación de texto
- Traducción de texto
- Resumen de texto
- Clasificación de texto
- Evaluación de calidad de texto
- Texto para codificar
- Reconocimiento de entidad nombrado
- Extracción de relación
- Multitarea
Conjuntos de datos de modelos de idiomas grandes (MLLMS) multimodales
- Corpus para capacitar
- Instrucciones de datos de ajuste fino
- Conjuntos de datos de evaluación
Conjuntos de datos de generación aumentada de recuperación (trapo)

Corpus para capacitar

Los corpus previos al entrenamiento son grandes colecciones de datos de texto utilizados durante el proceso de pre-entrenamiento de LLM.

Corporativos de pre-entrenamiento general

Los corpus generales previos a la capacitación son conjuntos de datos a gran escala compuestos por texto extenso de diversos dominios y fuentes. Su característica principal es que el contenido de texto no se limita a un solo dominio, lo que los hace más adecuados para entrenar modelos fundamentales generales. Los corpus se clasifican según las categorías de datos.

Formato de información del conjunto de datos:

 - Dataset name  Release Time | Public or Not | Language | Construction Method | Paper | Github | Dataset | Website
  - Publisher:
  - Size:
  - License:
  - Source:

Página web

CC-Stories 2018-6 | No | Es | CI | Papel | GitHub | Conjunto de datos
- Editorial: Google Brain
- Tamaño: 31 GB
- Licencia: -
- Fuente: Crawl común
CC100 2020-7 | Todos | Multi (100) | CI | Papel | Conjunto de datos
- Editorial: Facebook AI
- Tamaño: 2.5 TB
- Licencia: Términos de uso comunes de rastreo
- Fuente: Crawl común
Cluecorpus2020 2020-3 | Todos | Zh | CI | Papel | Conjunto de datos
- Editorial: organización de pista
- Tamaño: 100 GB
- Licencia: MIT
- Fuente: Crawl común
Crawl común 2007-x | Todos | Multi | Hg | Sitio web
- Editorial: Crawl común
- Tamaño: -
- Licencia: Términos de uso comunes de rastreo
- Fuente: datos de rastreadores web
Culturax 2023-9 | Todos | Multi (167) | CI | Papel | Conjunto de datos
- Editorial: Universidad de Oregon et al.
- Tamaño: 27 TB
- Licencia: Licencia MC4 y Oscar
- Fuente: MC4, Oscar
C4 2019-10 | Todos | Es | CI | Papel | Conjunto de datos
- Editorial: Google Research
- Tamaño: 12.68 TB
- Licencia: ODC-by y términos de uso comunes de rastreo
- Fuente: Crawl común
MC4 2021-6 | Todos | Multi (108) | CI | Papel | Conjunto de datos
- Editorial: Google Research
- Tamaño: 251 GB
- Licencia: ODC-by y términos de uso comunes de rastreo
- Fuente: Crawl común
Oscar 22.01 2022-1 | Todos | Multi (151) | CI | Papel | Conjunto de datos | Sitio web
- Editorial: Inria
- Tamaño: 8.41 TB
- Licencia: CC0
- Fuente: Crawl común
RealNews 2019-5 | Todos | Es | CI | Papel | Github
- Editorial: Universidad de Washington et al.
- Tamaño: 120 GB
- Licencia: Apache-2.0
- Fuente: Crawl común
Redpajama-V2 2023-10 | Todos | Multi (5) | CI | GitHub | Conjunto de datos | Sitio web
- Editor: Computadora juntos
- Tamaño: 30.4 t fichas
- Licencia: Términos de uso comunes de rastreo
- Fuente: Crawl común, C4, etc.
Refinedweb 2023-6 | Parcial | Es | CI | Papel | Conjunto de datos
- Editorial: el equipo Falcon LLM
- Tamaño: 5000 GB
- Licencia: ODC por 1.0
- Fuente: Crawl común
Wudaocorpora-text 2021-6 | Parcial | Zh | Hg | Papel | Conjunto de datos
- Editorial: Baai et al.
- Tamaño: 200 GB
- Licencia: MIT
- Fuente: Páginas web chinas
Wanjuan-CC 2024-2 | Parcial | Es | Hg | Papel | Conjunto de datos
- Editorial: Laboratorio de inteligencia artificial de Shanghai
- Tamaño: 1 T Tokens
- Licencia: CC-By-4.0
- Fuente: Crawl común
Madlad-400 2023-9 | Todos | Multi (419) | Hg | Papel | GitHub | Conjunto de datos
- Editorial: Google Deepmind et al.
- Tamaño: 2.8 t fichas
- Licencia: ODL-By
- Fuente: Crawl común
Fineweb 2024-4 | Todos | Es | CI | Conjunto de datos
- Editorial: HuggingfaceFW
- Tamaño: tokens de 15 tb
- Licencia: ODC por 1.0
- Fuente: Crawl común
CCI 2.0 2024-4 | Todos | Zh | Hg | DataSet1 | Conjunto de datos2
- Editorial: Baai
- Tamaño: 501 GB
- Licencia: Agresión de uso de CCI
- Fuente: Páginas web chinas
DCLM 2024-6 | Todos | Es | CI | Papel | GitHub | Conjunto de datos | Sitio web
- Editorial: Universidad de Washington et al.
- Tamaño: 279.6 TB
- Licencia: Términos de uso comunes de rastreo
- Fuente: Crawl común

Textos del idioma

ANC 2003-X | Todos | Es | Hg | Sitio web
- Editorial: La Fundación Nacional de Ciencias de los Estados Unidos et al.
- Tamaño: -
- Licencia: -
- Fuente: Textos en inglés americano
BNC 1994-X | Todos | Es | Hg | Sitio web
- Editorial: Oxford University Press et al.
- Tamaño: 4124 textos
- Licencia: -
- Fuente: Textos en inglés británico
News-Rrawl 2019-1 | Todos | Multi (59) | Hg | Conjunto de datos
- Editorial: Ucri et al.
- Tamaño: 110 GB
- Licencia: CC0
- Fuente: periódicos

Libros

Archivo de Anna 2023-X | Todos | Multi | Hg | Sitio web
- Editorial: Anna
- Tamaño: 586.3 TB
- Licencia: -
- Fuente: Sci-Hub, Biblioteca Genesis, Z-Biblioteca, etc.
BookCorpusopen 2021-5 | Todos | Es | CI | Papel | GitHub | Conjunto de datos
- Editorial: Jack Bandy et al.
- Tamaño: 17,868 libros
- Licencia: Smashwords Términos de servicio
- Fuente: Toronto Book Corpus
PG-19 2019-11 | Todos | Es | Hg | Papel | GitHub | Conjunto de datos
- Editorial: DeepMind
- Tamaño: 11.74 GB
- Licencia: Apache-2.0
- Fuente: Proyecto Gutenberg
Proyecto Gutenberg 1971-X | Todos | Multi | Hg | Sitio web
- Editorial: Ibiblio et al.
- Tamaño: -
- Licencia: El Proyecto Gutenberg
- Fuente: datos de libros electrónicos
Smashwords 2008-X | Todos | Multi | Hg | Sitio web
- Editor: Draft2Digital et al.
- Tamaño: -
- Licencia: Smashwords Términos de servicio
- Fuente: datos de libros electrónicos
Toronto Book Corpus 2015-6 | No | Es | Hg | Papel | Sitio web
- Editorial: Universidad de Toronto et al.
- Tamaño: 11,038 libros
- Licencia: MIT & Smashwords Términos de servicio
- Fuente: Smashwords

Materiales académicos

ARXIV 1991-X | Todos | Es | Hg | Sitio web
- Editorial: Paul Ginsparg et al.
- Tamaño: -
- Licencia: Términos de uso para API ARXIV
- Fuente: preimpresión de Arxiv
S2ORC 2020-6 | Todos | Es | CI | Papel | Github
- Editorial: Ai2 et al.
- Tamaño: 81.1 MB
- Licencia: ODC por 1.0
- Fuente: Mag, Arxiv, PubMed, etc.

Código

BigQuery 2022-3 | No | PL | CI | Papel | Github
- Editorial: Salesforce Research
- Tamaño: 341.1 GB
- Licencia: Apache-2.0
- Fuente: BigQuery
Github 2008-4 | Todos | PL | Hg | Sitio web
- Editorial: Microsoft
- Tamaño: -
- Licencia: -
- Fuente: Varios proyectos de código
PHI-1 2023-6 | No | EN & PL | HG & MC | Papel | Conjunto de datos
- Editorial: Microsoft Research
- Tamaño: 7 B Tokens
- Licencia: CC-By-NC-SA-3.0
- Fuente: The Stack, StackOverflow, GPT-3.5 Generación
La pila 2022-11 | Todos | PL (358) | Hg | Papel | Conjunto de datos
- Editorial: ServiceNow Research et al.
- Tamaño: 6 TB
- Licencia: los términos de las licencias originales
- Fuente: archivos de código fuente con licencia permisiva

Corpus paralelo

MTP 2023-9 | Todos | En & zh | HG & CI | Conjunto de datos
- Editorial: Baai
- Tamaño: 1.3 TB
- Licencia: Protocolo de uso de datos BAAI
- Fuente: Pares de texto paralelos chinos-inglés en la web
Multiun 2010-5 | Todos | Multi (7) | Hg | Papel | Sitio web
- Editorial: Centro de Investigación Alemán de Inteligencia Artificial (DFKI) GMBH
- Tamaño: 4353 MB
- Licencia: -
- Fuente: Documentos de las Naciones Unidas
Paracrawl 2020-7 | Todos | Multi (42) | Hg | Papel | Sitio web
- Editorial: Prompsit et al.
- Tamaño: 59996 archivos
- Licencia: CC0
- Fuente: datos de rastreadores web
Uncorpus v1.0 2016-5 | Todos | Multi (6) | Hg | Papel | Sitio web
- Editorial: Naciones Unidas et al.
- Tamaño: 799276 archivos
- Licencia: -
- Fuente: Documentos de las Naciones Unidas
Caligned 2020-11 | Todos | Multi (138) | Hg | Papel | Conjunto de datos
- Editorial: Facebook Ai et al.
- Tamaño: pares de URL de 392 m
- Licencia: -
- Fuente: Crawl común
Wikimatrix 2021-4 | Todos | Multi (85) | Hg | Papel | GitHub | Conjunto de datos
- Editorial: Facebook Ai et al.
- Tamaño: 134 m oraciones paralelas
- Licencia: CC-by-SA
- Fuente: Wikipedia

Redes sociales

OpenWebText 2019-4 | Todos | Es | Hg | Sitio web
- Editorial: Universidad de Brown
- Tamaño: 38 GB
- Licencia: CC0
- Fuente: Reddit
PushShift Reddit 2020-1 | Todos | Es | CI | Papel | Sitio web
- Editorial: Pushshift.io et al.
- Tamaño: 2 TB
- Licencia: -
- Fuente: Reddit
Reddit 2005-6 | Todos | Es | Hg | Sitio web
- Editorial: Condé Nast Digital et al.
- Tamaño: -
- Licencia: -
- Fuente: Publicaciones en las redes sociales
Stackexchange 2008-9 | Todos | Es | Hg | Conjunto de datos | Sitio web
- Editorial: Stack Exchange
- Tamaño: -
- Licencia: CC-By-SA-4.0
- Fuente: Datos de preguntas y respuestas de la comunidad
WebText 2019-2 | Parcial | Es | Hg | Papel | GitHub | Conjunto de datos
- Editorial: OpenAI
- Tamaño: 40 GB
- Licencia: MIT
- Fuente: Reddit
Zhihu 2011-1 | Todos | Zh | Hg | Sitio web
- Editorial: Beijing Zhizhe Tianxia Technology Co., Ltd
- Tamaño: -
- Licencia: Acuerdo de usuario de Zhihu
- Fuente: Publicaciones en las redes sociales

Enciclopedia

Baidu Baike 2008-4 | Todos | Zh | Hg | Sitio web
- Editorial: Baidu
- Tamaño: -
- Licencia: Acuerdo de usuario de Baidu Baike
- Fuente: datos de contenido enciclopédico
Tigerbot-Wiki 2023-5 | Todos | Zh | Hg | Papel | GitHub | Conjunto de datos
- Editorial: Tigerbot
- Tamaño: 205 MB
- Licencia: Apache-2.0
- Fuente: Baidu Baike
Wikipedia 2001-1 | Todos | Multi | Hg | Conjunto de datos | Sitio web
- Editorial: Fundación Wikimedia
- Tamaño: -
- Licencia: CC-BY-SA-3.0 y GFDL
- Fuente: datos de contenido enciclopédico

Multicatategoría

Arabictext 2022 2022-12 | Todos | AR | HG & CI | Conjunto de datos
- Editorial: Baai et al.
- Tamaño: 201.9 GB
- Licencia: CC-By-SA-4.0
- Fuente: Arabicweb, Oscar, CC100, etc.
MNBVC 2023-1 | Todos | Zh | HG & CI | GitHub | Conjunto de datos
- Editorial: Comunidad de Liwu
- Tamaño: 20811 GB
- Licencia: MIT
- Fuente: Libros chinos, páginas web, tesis, etc.
Redpajama-V1 2023-4 | Todos | Multi | HG & CI | GitHub | Conjunto de datos
- Editor: Computadora juntos
- Tamaño: 1.2 t fichas
- Licencia: -
- Fuente: Crawl común, Github, libros, etc.
Roots 2023-3 | Parcial | Multi (59) | HG & CI | Papel | Conjunto de datos
- Editorial: Abrazando a Face et al.
- Tamaño: 1.61 TB
- Licencia: Bloom Open-Rail-M
- Fuente: Oscar, Github, etc.
La pila 2021-1 | Todos | Es | HG & CI | Papel | GitHub | Conjunto de datos
- Editorial: Eleutherai
- Tamaño: 825.18 GB
- Licencia: MIT
- Fuente: Libros, Arxiv, Github, etc.
Tigerbot_pretrain_en 2023-5 | Parcial | Es | CI | Papel | GitHub | Conjunto de datos
- Editorial: Tigerbot
- Tamaño: 51 GB
- Licencia: Apache-2.0
- Fuente: libros en inglés, páginas web, en-wiki, etc.
Tigerbot_pretrain_zh 2023-5 | Parcial | Zh | Hg | Papel | GitHub | Conjunto de datos
- Editorial: Tigerbot
- Tamaño: 55 GB
- Licencia: Apache-2.0
- Fuente: Libros chinos, páginas web, zh-wiki, etc.
Wanjuantext-1.0 2023-8 | Todos | Zh | Hg | Papel | GitHub | Conjunto de datos
- Editorial: Laboratorio de Shanghai AI
- Tamaño: 1094 GB
- Licencia: CC-By-4.0
- Fuente: Páginas web, enciclopedia, libros, etc.
Dolma 2024-1 | Todos | Es | HG & CI | Papel | GitHub | Conjunto de datos
- Editorial: Ai2 et al.
- Tamaño: 11519 GB
- Licencia: MR Acuerdo
- Fuente: Proyecto Gutenberg, C4, Reddit, etc.
Slimpajama 2023-6 | Todos | Es | HG & CI | GitHub | Conjunto de datos | Sitio web
- Editorial: Cerebras et al.
- Tamaño: 627 B Tokens
- Licencia: -
- Fuente: Crawl común, C4, Github, etc.
Massivetext 2021-12 | No | Multi | HG & CI | Papel
- Editorial: Google Deepmind
- Tamaño: 10.5 TB
- Licencia: -
- Fuente: MassiveWeb, C4, libros, etc.
Minerva 2022-6 | No | Es | Hg | Papel
- Editorial: Google Research
- Tamaño: 38.5 B Tokens
- Licencia: -
- Fuente: ARXIV, Páginas web, etc.
MAP-CC 2024-4 | Todos | Zh | Hg | Papel | GitHub | Conjunto de datos | Sitio web
- Editorial: Comunidad de investigación de proyección de arte multimodal et al.
- Tamaño: 840.48 B Token
- Licencia: CC-By-NC-ND-4.0
- Fuente: Crawl común chino, enciclopedias chinas, libros chinos, etc.
Expositoria-Prose-V1 2024-8 | Todos | Es | HG & CI | Papel | GitHub | Conjunto de datos
- Editorial: Pints.ai Labs
- Tamaño: 56 B Tokens
- Licencia: MIT
- Fuente: Arxiv, Wikipedia, Gutenberg, etc.

Corpus de pre-entrenamiento específicos del dominio

Los corpus de pre-entrenamiento específicos del dominio son conjuntos de datos LLM personalizados para campos o temas específicos. El tipo de corpus se emplea típicamente en la fase incremental de pre-entrenamiento de LLM. Los corpus se clasifican según los dominios de datos.

Formato de información del conjunto de datos:

 - Dataset name  Release Time | Public or Not | Language | Construction Method | Paper | Github | Dataset | Website
  - Publisher:
  - Size:
  - License:
  - Source:
  - Category:
  - Domain:

Financiero

BBT-Fincorpus 2023-2 | Parcial | Zh | Hg | Papel | GitHub | Sitio web
- Editorial: Fudan University et al.
- Tamaño: 256 GB
- Licencia: -
- Fuente: Anuncios de la empresa, informes de investigación, financiero
- Categoría: Multi
- Dominio: Finanzas
Fincorpus 2023-9 | Todos | Zh | Hg | Papel | GitHub | Conjunto de datos
- Editorial: Du Xiaoman
- Tamaño: 60.36 GB
- Licencia: Apache-2.0
- Fuente: Anuncios de la empresa, noticias financieras, preguntas del examen financiero
- Categoría: Multi
- Dominio: Finanzas
Finglm 2023-7 | Todos | Zh | Hg | Github
- Editorial: Conocimiento Atlas et al.
- Tamaño: 69 GB
- Licencia: Apache-2.0
- Fuente: Informes anuales de empresas cotizadas
- Categoría: Textos del idioma
- Dominio: Finanzas
Tigerbot-gana 2023-5 | Todos | Zh | Hg | Papel | GitHub | Conjunto de datos
- Editorial: Tigerbot
- Tamaño: 488 MB
- Licencia: Apache-2.0
- Fuente: Informes financieros
- Categoría: Textos del idioma
- Dominio: Finanzas
Tigerbot-Research 2023-5 | Todos | Zh | Hg | Papel | GitHub | Conjunto de datos
- Editorial: Tigerbot
- Tamaño: 696 MB
- Licencia: Apache-2.0
- Fuente: Informes de investigación
- Categoría: Textos del idioma
- Dominio: Finanzas

Médico

Medical-Pt 2023-5 | Todos | Zh | CI | GitHub | Conjunto de datos
- Editorial: Ming Xu
- Tamaño: 632.78 MB
- Licencia: Apache-2.0
- Fuente: Datos de enciclopedia médica, libros de texto médicos
- Categoría: Multi
- Dominio: médico
PubMed Central 2000-2 | Todos | Es | Hg | Sitio web
- Editorial: NCBI
- Tamaño: -
- Licencia: Aviso de derechos de autor de PMC
- Fuente: literatura biomédica científica
- Categoría: materiales académicos
- Dominio: médico

Matemáticas

Prueba-2 2023-10 | Todos | Es | HG & CI | Papel | GitHub | Conjunto de datos | Sitio web
- Editorial: Princeton University et al.
- Tamaño: 55 B Tokens
- Licencia: -
- Fuente: Arxiv, OpenWebMath, Algebraicstack
- Categoría: Multi
- Dominio: Matemáticas
Mathpile 2023-12 | Todos | Es | Hg | Papel | GitHub | Conjunto de datos
- Editorial: Shanghai Jiao Tong University et al.
- Tamaño: tokens 9.5 B
- Licencia: CC-By-NC-SA-4.0
- Fuente: Libros de texto, Wikipedia, Proofwiki, CommonCrawl, stackexchange, arxiv
- Categoría: Multi
- Dominio: Matemáticas
OpenWebmath 2023-10 | Todos | Es | Hg | Papel | GitHub | Conjunto de datos
- Editorial: Universidad de Toronto et al.
- Tamaño: 14.7 B Tokens
- Licencia: ODC por 1.0
- Fuente: Crawl común
- Categoría: páginas web
- Dominio: Matemáticas

Otro

Tigerbot-Law 2023-5 | Todos | Zh | Hg | Papel | GitHub | Conjunto de datos
- Editorial: Tigerbot
- Tamaño: 29.9 MB
- Licencia: Apache-2.0
- Fuente: cláusulas legales
- Categoría: Textos del idioma
- Dominio: Ley
Transgpt-Pt 2023-7 | Todos | Zh | Hg | GitHub | Conjunto de datos
- Editorial: Universidad de Beijing Jiaotong
- Tamaño: 35.8 MB
- Licencia: Apache-2.0
- Fuente: Documentos tecnológicos, información de construcción de ingeniería, datos estadísticos, etc.
- Categoría: Multi
- Dominio: transporte

Instrucciones de datos de ajuste fino

Los conjuntos de datos de ajuste de instrucción consisten en una serie de pares de texto que comprenden "entradas de instrucciones" y "salidas de respuesta". Las "entradas de instrucciones" representan solicitudes hechas por humanos al modelo. Existen varios tipos de instrucciones, como clasificación, resumen, paráfrasis, etc. "Respuestas salidas" son las respuestas generadas por el modelo que siguen la instrucción y se alinean con las expectativas humanas.

Instrucción general para conjuntos de datos ajustados

Instrucción general Los conjuntos de datos de ajuste fino contienen una o más categorías de instrucciones sin restricciones de dominio, principalmente con el objetivo de mejorar la capacidad de seguimiento de instrucciones de LLM en tareas generales. Los conjuntos de datos se clasifican según los métodos de construcción.

Formato de información del conjunto de datos:

 - Dataset name  Release Time | Public or Not | Language | Construction Method | Paper | Github | Dataset | Website
  - Publisher:
  - Size:
  - License:
  - Source:
  - Instruction Category:

Conjuntos de datos generados por humanos (HG)

databricks-dolly-15k 2023-4 | Todos | Es | Hg | Conjunto de datos | Sitio web
- Editorial: Databricks
- Tamaño: 15011 instancias
- Licencia: CC-By-SA-3.0
- Fuente: Generado manualmente en función de diferentes categorías de instrucciones
- Categoría de instrucciones: multi
InstructionWild_V2 2023-6 | Todos | En & zh | Hg | Github
- Editorial: Universidad Nacional de Singapur
- Tamaño: 110k instancias
- Licencia: -
- Fuente: recopilado en la web
- Categoría de instrucciones: multi
LCCC 2020-8 | Todos | Zh | Hg | Papel | Github
- Editorial: Tsinghua University et al.
- Tamaño: 12m instancias
- Licencia: MIT
- Fuente: Interacciones de usuario de rastreo en las redes sociales
- Categoría de instrucciones: multi
OASST1 2023-4 | Todos | Multi (35) | Hg | Papel | GitHub | Conjunto de datos
- Editorial: OpenSistant
- Tamaño: 161443 instancias
- Licencia: Apache-2.0
- Fuente: Generado y anotado por humanos
- Categoría de instrucciones: multi
OL-CC 2023-6 | Todos | Zh | Hg | Conjunto de datos
- Editorial: Baai
- Tamaño: 11655 instancias
- Licencia: Apache-2.0
- Fuente: Generado y anotado por humanos
- Categoría de instrucciones: multi
Zhihu-kol 2023-3 | Todos | Zh | Hg | GitHub | Conjunto de datos
- Editorial: Wangrui6
- Tamaño: 1006218 instancias
- Licencia: MIT
- Fuente: Rastrear desde Zhihu
- Categoría de instrucciones: multi
AYA DataSet 2024-2 | Todos | Multi (65) | Hg | Papel | Conjunto de datos | Sitio web
- Editorial: Cohere for AI Community et al.
- Tamaño: 204k instancias
- Licencia: Apache-2.0
- Fuente: recopilado y anotado manualmente a través de la plataforma de anotación de AYA
- Categoría de instrucciones: multi
Instructora 2023-5 | Todos | En & zh | Hg | Papel | GitHub | Conjunto de datos
- Editorial: Zhejiang University et al.
- Tamaño: 371700 instancias
- Licencia: MIT
- Fuente: Baidu Baike, Wikipedia
- Categoría de instrucciones: extracción

Conjuntos de datos construidos con modelo (MC)

Alpaca_data 2023-3 | Todos | Es | MC | Github
- Editorial: Stanford Alpaca
- Tamaño: 52k instancias
- Licencia: Apache-2.0
- Fuente: Generado por Text-Davinci-003 con indicaciones de APLACA_DATA
- Categoría de instrucciones: multi
Belle_generated_chat 2023-5 | Todos | Zh | MC | GitHub | Conjunto de datos
- Editorial: Belle
- Tamaño: 396004 instancias
- Licencia: GPL-3.0
- Fuente: Generado por chatgpt
- Categoría de instrucciones: generación
Belle_multiturn_chat 2023-5 | Todos | Zh | MC | GitHub | Conjunto de datos
- Editorial: Belle
- Tamaño: 831036 instancias
- Licencia: GPL-3.0
- Fuente: Generado por chatgpt
- Categoría de instrucciones: multi
Belle_train_0.5m_cn 2023-4 | Todos | Zh | MC | GitHub | Conjunto de datos
- Editorial: Belle
- Tamaño: 519255 instancias
- Licencia: GPL-3.0
- Fuente: Generado por Text-Davinci-003
- Categoría de instrucciones: multi
Belle_train_1m_cn 2023-4 | Todos | Zh | MC | GitHub | Conjunto de datos
- Editorial: Belle
- Tamaño: 917424 instancias
- Licencia: GPL-3.0
- Fuente: Generado por Text-Davinci-003
- Categoría de instrucciones: multi
Belle_train_2m_cn 2023-5 | Todos | Zh | MC | GitHub | Conjunto de datos
- Editorial: Belle
- Tamaño: 2m instancias
- Licencia: GPL-3.0
- Fuente: Generado por chatgpt
- Categoría de instrucciones: multi
Belle_train_3.5m_cn 2023-5 | Todos | Zh | MC | GitHub | Conjunto de datos
- Editorial: Belle
- Tamaño: 3606402 instancias
- Licencia: GPL-3.0
- Fuente: Generado por chatgpt
- Categoría de instrucciones: multi
Camel 2023-3 | Todos | Multi & PL | MC | Papel | GitHub | Conjunto de datos | Sitio web
- Editorial: KAUST
- Tamaño: 1659328 instancias
- Licencia: CC-By-NC-4.0
- Fuente: Diálogo generado por dos agentes GPT-3.5-Turbo
- Categoría de instrucciones: multi
Chatgpt_corpus 2023-6 | Todos | Zh | MC | Github
- Editorial: Plexpt
- Tamaño: 3270k instancias
- Licencia: GPL-3.0
- Fuente: Generado por GPT-3.5-TURBO
- Categoría de instrucciones: multi
InstructionWild_V1 2023-3 | Todos | En & zh | MC | Github
- Editorial: Universidad Nacional de Singapur
- Tamaño: 104k instancias
- Licencia: -
- Fuente: Generado por Operai API
- Categoría de instrucciones: multi
LMSYS-CHAT-1M 2023-9 | Todos | Multi | MC | Papel | Conjunto de datos
- Editorial: UC Berkeley et al.
- Tamaño: 1M Instancias
- Licencia: Licencia LMSYS-CHAT-1M
- Fuente: Generado por múltiples LLMS
- Categoría de instrucciones: multi
Moss_002_sft_data 2023-4 | Todos | En & zh | MC | GitHub | Conjunto de datos
- Editorial: Universidad de Fudan
- Tamaño: 1161137 instancias
- Licencia: CC-By-NC-4.0
- Fuente: Generado por Text-Davinci-003
- Categoría de instrucciones: multi
Moss_003_sft_data 2023-4 | Todos | En & zh | MC | GitHub | Conjunto de datos
- Editorial: Universidad de Fudan
- Tamaño: 1074551 instancias
- Licencia: CC-By-NC-4.0
- Fuente: Datos de conversación de MOSS-002 y generados por GPT-3.5-TURBO
- Categoría de instrucciones: multi
Moss_003_sft_plugin_data 2023-4 | Parcial | En & zh | MC | GitHub | Conjunto de datos
- Editorial: Universidad de Fudan
- Tamaño: 300k instancias
- Licencia: CC-By-NC-4.0
- Fuente: Generado por Plugins y LLMS
- Categoría de instrucciones: multi
OpenChat 2023-7 | Todos | Es | MC | Papel | GitHub | Conjunto de datos
- Editorial: Tsinghua University et al.
- Tamaño: 70k instancias
- Licencia: MIT
- Fuente: ShareGPT
- Categoría de instrucciones: multi
REDGPT-DATASET-V1-CN 2023-4 | Parcial | Zh | MC | Github
- Editorial: Da-Southampton
- Tamaño: 50k instancias
- Licencia: Apache-2.0
- Fuente: Generado por LLMS
- Categoría de instrucciones: multi
Autoinstructo 2022-12 | Todos | Es | MC | Papel | Github
- Editorial: Universidad de Washington et al.
- Tamaño: 52445 instancias
- Licencia: Apache-2.0
- Fuente: Generado por GPT-3
- Categoría de instrucciones: multi
Sharechat 2023-4 | Todos | Multi | MC | Sitio web
- Editor: Sharechat
- Tamaño: 90k instancias
- Licencia: CC0
- Fuente: ShareGPT
- Categoría de instrucciones: multi
ShareGPT-Chinese-English-90k 2023-7 | Todos | En & zh | MC | GitHub | Conjunto de datos
- Editor: Shareai
- Tamaño: 90k instancias
- Licencia: Apache-2.0
- Fuente: ShareGPT
- Categoría de instrucciones: multi
ShareGPT90K 2023-4 | Todos | Es | MC | Conjunto de datos
- Editorial: Ryokoai
- Tamaño: 90k instancias
- Licencia: CC0
- Fuente: ShareGPT
- Categoría de instrucciones: multi
Ultrachat 2023-5 | Todos | Es | MC | Papel | Github
- Editorial: Universidad de Tsinghua
- Tamaño: 1468352 instancias
- Licencia: CC-By-NC-4.0
- Fuente: Diálogo generado por dos agentes de Chatgpt
- Categoría de instrucciones: multi
Instrucciones antinaturales 2022-12 | Todos | Es | MC | Papel | Github
- Editorial: Tel Aviv University et al.
- Tamaño: 240670 instancias
- Licencia: MIT
- Fuente: Generado por LLMS
- Categoría de instrucciones: multi
WebGLM-QA 2023-6 | Todos | Es | MC | Papel | GitHub | Conjunto de datos
- Editorial: Tsinghua University et al.
- Tamaño: 44979 instancias
- Licencia: Apache-2.0
- Fuente: Construct WebGLM-QA a través de LLM In-Context Bootstrapping
- Categoría de instrucciones: Open QA
Wizard_evol_instruct_196k 2023-6 | Todos | Es | MC | Papel | GitHub | Conjunto de datos
- Editorial: Microsoft et al.
- Tamaño: 196k instancias
- Licencia: -
- Fuente: Evolucionar las instrucciones a través del método Evol-Instructo
- Categoría de instrucciones: multi
Wizard_evol_instruct_70k 2023-5 | Todos | Es | MC | Papel | GitHub | Conjunto de datos
- Editorial: Microsoft et al.
- Tamaño: 70k instancias
- Licencia: -
- Fuente: Evolucionar las instrucciones a través del método Evol-Instructo
- Categoría de instrucciones: multi
Wildchat 2024-5 | Parcial | Multi | MC | Papel | Conjunto de datos
- Editorial: Cornell University et al.
- Tamaño: 1039785 instancias
- Licencia: Licencia de impacto AI2
- Fuente: Conversaciones entre usuarios y chatgpt, GPT-4
- Categoría de instrucciones: multi

Colección y mejora de los conjuntos de datos existentes (CI)

CrossFit 2021-4 | Todos | Es | CI | Papel | Github
- Editorial: Universidad del Sur de California
- Tamaño: 269 conjuntos de datos
- Licencia: -
- Fuente: Colección y mejora de varios conjuntos de datos de PNL
- Categoría de instrucciones: multi
DialogStudio 2023-7 | Todos | Es | CI | Papel | GitHub | Conjunto de datos
- Editorial: Salesforce Ai et al.
- Tamaño: 87 conjuntos de datos
- Licencia: Apache-2.0
- Fuente: Colección y mejora de varios conjuntos de datos de PNL
- Categoría de instrucciones: multi
Dynosaur 2023-5 | Todos | Es | CI | Papel | GitHub | Conjunto de datos | Sitio web
- Editorial: UCLA et al.
- Tamaño: 801900 instancias
- Licencia: Apache-2.0
- Fuente: Colección y mejora de varios conjuntos de datos de PNL
- Categoría de instrucciones: multi
Flan-Mini 2023-7 | Todos | Es | CI | Papel | GitHub | Conjunto de datos
- Editorial: Universidad Tecnológica y Diseño de Singapur
- Tamaño: 1.34m instancias
- Licencia: CC
- Fuente: Recopilación y mejora de varios conjuntos de datos de ajuste de instrucciones
- Categoría de instrucciones: multi
Flan 2021 2021-9 | Todos | Multi | CI | Papel | Github
- Editorial: Google Research
- Tamaño: 62 conjuntos de datos
- Licencia: Apache-2.0
- Fuente: Colección y mejora de varios conjuntos de datos de PNL
- Categoría de instrucciones: multi
FLAN 2022 2023-1 | Parcial | Multi | CI | Papel | GitHub | Conjunto de datos
- Editorial: Google Research
- Tamaño: 1836 conjuntos de datos
- Licencia: Apache-2.0
- Fuente: Recopilación y mejora de varios conjuntos de datos de ajuste de instrucciones
- Categoría de instrucciones: multi
Instrucciones 2022-5 | Todos | Es | CI | Papel | Github
- Editorial: Universidad Carnegie Mellon
- Tamaño: 59 conjuntos de datos
- Licencia: Apache-2.0
- Fuente: Colección y mejora de varios conjuntos de datos de PNL
- Categoría de instrucciones: multi
Instrucciones naturales 2021-4 | Todos | Es | CI | Papel | GitHub | Conjunto de datos
- Editorial: Allen Institute for AI et al.
- Tamaño: 61 conjuntos de datos
- Licencia: Apache-2.0
- Fuente: Colección y mejora de varios conjuntos de datos de PNL
- Categoría de instrucciones: multi
OIG 2023-3 | Todos | Es | CI | Conjunto de datos
- Editorial: Laion
- Tamaño: 3878622 instancias
- Licencia: Apache-2.0
- Fuente: Colección y mejora de varios conjuntos de datos
- Categoría de instrucciones: multi
Open-Platypus 2023-8 | Todos | Es | CI | Papel | GitHub | Conjunto de datos | Sitio web
- Editorial: Boston University
- Tamaño: 24926 instancias
- Licencia: -
- Fuente: Colección y mejora de varios conjuntos de datos
- Categoría de instrucciones: multi
OPT-IML BENCH 2022-12 | No | Multi | CI | Papel | Github
- Editor: meta ai
- Tamaño: 2000 conjuntos de datos
- Licencia: MIT
- Fuente: Colección y mejora de varios conjuntos de datos de PNL
- Categoría de instrucciones: multi
PromotSource 2022-2 | Todos | Es | CI | Papel | Github
- Editorial: Brown University et al.
- Tamaño: 176 conjuntos de datos
- Licencia: Apache-2.0
- Fuente: Colección y mejora de varios conjuntos de datos de PNL
- Categoría de instrucciones: multi
Instrucciones súper naturales 2022-4 | Todos | Multi | CI | Papel | Github
- Editor: Univ. de Washington et al.
- Tamaño: 1616 conjuntos de datos
- Licencia: Apache-2.0
- Fuente: Colección y mejora de varios conjuntos de datos de PNL
- Categoría de instrucciones: multi
T0 2021-10 | Todos | Es | CI | Papel | DataSet1 | Conjunto de datos2
- Publisher: Hugging Face et al.
- Size: 62 datasets
- License: Apache-2.0
- Source: Collection and improvement of various NLP datasets
- Instruction Category: Multi
UnifiedSKG 2022-3 | All | EN | CI | Papel | Github
- Publisher: The University of Hong Kong et al.
- Size: 21 datasets
- License: Apache-2.0
- Source: Collection and improvement of various NLP datasets
- Instruction Category: Multi
xP3 2022-11 | All | Multi (46) | CI | Papel | Github
- Publisher: Hugging Face et al.
- Size: 82 datasets
- License: Apache-2.0
- Source: Collection and improvement of various NLP datasets
- Instruction Category: Multi
IEPile 2024-2 | All | EN & ZH | CI | Papel | Github | Conjunto de datos
- Publisher: Zhejiang University et al.
- Size: 33 datasets
- License: CC-BY-NC-SA-4.0
- Source: Collection and improvement of various IE datasets
- Instruction Category: Extraction
KOLLM-Conversations 2024-3 | All | KO | CI | Conjunto de datos
- Publisher: davidkim205
- Size: 1122566 instances
- License: Apache-2.0
- Source: Collection and improvement of Korean datasets
- Instruction Category: Multi

HG & CI

Firefly 2023-4 | All | ZH | HG & CI | Github | Conjunto de datos
- Publisher: YeungNLP
- Size: 1649399 instances
- License: -
- Source: Collect Chinese NLP datasets and manually generate data related to Chinese culture
- Instruction Category: Multi
LIMA-sft 2023-5 | All | EN | HG & CI | Papel | Conjunto de datos
- Publisher: Meta AI et al.
- Size: 1330 instances
- License: CC-BY-NC-SA
- Source: Manually select from various types of data
- Instruction Category: Multi
COIG-CQIA 2024-3 | All | ZH | HG & CI | Papel | Conjunto de datos
- Publisher: Shenzhen Institute of Advanced Technology et al.
- Size: 48375 instances
- License: -
- Source: Q&A communities, Wikipedia, examinations, existing NLP datasets
- Instruction Category: Multi

HG & MC

InstructGPT-sft 2022-3 | Not | EN | HG & MC | Papel
- Publisher: OpenAI
- Size: 14378 instances
- License: -
- Source: Platform Q&A data and manual labeling
- Instruction Category: Multi

CI & MC

Alpaca_GPT4_data 2023-4 | All | EN | CI & MC | Papel | Github
- Publisher: Microsoft Research
- Size: 52K instances
- License: Apache-2.0
- Source: Generated by GPT-4 with Aplaca_data prompts
- Instruction Category: Multi
Alpaca_GPT4_data_zh 2023-4 | All | ZH | CI & MC | Github | Conjunto de datos
- Publisher: Microsoft Research
- Size: 52K instances
- License: Apache-2.0
- Source: Generated by GPT-4 with Alpaca_data prompts translated into Chinese by ChatGPT
- Instruction Category: Multi
Bactrain-X 2023-5 | All | Multi (52) | CI & MC | Papel | Github | Conjunto de datos
- Publisher: MBZUAI
- Size: 3484884 instances
- License: CC-BY-NC-4.0
- Source: Generated by GPT-3.5-Turbo with Aplaca_data and databricks-dolly-15K prompts translated into 51 languages by Google Translate API
- Instruction Category: Multi
Baize 2023-3 | Partial | EN | CI & MC | Papel | Github | Conjunto de datos
- Publisher: University of California et al.
- Size: 210311 instances
- License: GPL-3.0
- Source: Sample seeds from specific datasets to create multi-turn dialogues using ChatGPT
- Instruction Category: Multi
GPT4All 2023-3 | All | EN | CI & MC | Papel | Github | Conjunto de datos
- Publisher: nomic-ai
- Size: 739259 instances
- License: MIT
- Source: Generated by GPT-3.5-Turbo with other datasets' prompts
- Instruction Category: Multi
GuanacoDataset 2023-3 | All | Multi | CI & MC | Dataset | Sitio web
- Publisher: JosephusCheung
- Size: 534530 instances
- License: GPL-3.0
- Source: Expand upon the initial 52K dataset from the Alpaca model
- Instruction Category: Multi
LaMini-LM 2023-4 | All | EN | CI & MC | Papel | Github | Conjunto de datos
- Publisher: Monash University et al.
- Size: 2585615 instances
- License: CC-BY-NC-4.0
- Source: Generated by ChatGPT with synthetic and existing prompts
- Instruction Category: Multi
LogiCoT 2023-5 | All | EN & ZH | CI & MC | Papel | Github | Conjunto de datos
- Publisher: Westlake University et al.
- Size: 604840 instances
- License: CC-BY-NC-ND-4.0
- Source: Expand the datasets using GPT-4
- Instruction Category: Reasoning
LongForm 2023-4 | All | EN | CI & MC | Papel | Github | Conjunto de datos
- Publisher: LMU Munich et al.
- Size: 27739 instances
- License: MIT
- Source: Select documents from existing corpora and generating prompts for the documents using LLMs
- Instruction Category: Multi
Luotuo-QA-B 2023-5 | All | EN & ZH | CI & MC | Github | Conjunto de datos
- Publisher: Luotuo
- Size: 157320 instances
- License: Apache-2.0 & CC0
- Source: Use LLMs to generate Q&A pairs on CSL, arXiv, and CNN-DM datasets
- Instruction Category: Multi
OpenOrca 2023-6 | All | Multi | CI & MC | Papel | Conjunto de datos
- Publisher: Microsoft Researc
- Size: 4233923 instances
- License: MIT
- Source: Expand upon the Flan 2022 dataset using GPT-3.5-Turbo and GPT-4
- Instruction Category: Multi
Wizard_evol_instruct_zh 2023-5 | All | ZH | CI & MC | Github | Conjunto de datos
- Publisher: Central China Normal University et al.
- Size: 70K instances
- License: CC-BY-4.0
- Source: Generated by GPT with Wizard_evol_instruct prompts translated into Chinese
- Instruction Category: Multi
Lithuanian-QA-v1 2024-8 | All | LT | CI & MC | Papel | Conjunto de datos
- Publisher: Neurotechnology
- Size: 13848 instances
- License: CC-BY-4.0
- Source: Use ChatGPT to generate Q&A pairs on Wikipedia corpus
- Instruction Category: Multi
LongWriter-6K 2024-8 | All | EN & ZH | CI & MC | Papel | Github | Conjunto de datos
- Publisher: Tsinghua University et al.
- Size: 6000 instances
- License: Apache-2.0
- Source: Generated by GPT-4o with open-source datasets' prompts
- Instruction Category: Multi

HG & CI & MC

COIG 2023-4 | All | ZH | HG & CI & MC | Papel | Github | Conjunto de datos
- Publisher: BAAI
- Size: 191191 instances
- License: Apache-2.0
- Source: Translated instructions, Leetcode, Chinese exams, etc.
- Instruction Category: Multi
HC3 2023-1 | All | EN & ZH | HG & CI & MC | Papel | Github | Dataset1 | Dataset2
- Publisher: SimpleAI
- Size: 37175 instances
- License: CC-BY-SA-4.0
- Source: Human-Q&A pairs and ChatGPT-Q&A pairs from Q&A platforms, encyclopedias, etc.
- Instruction Category: Multi
Phoenix-sft-data-v1 2023-5 | All | Multi | HG & CI & MC | Papel | Github | Conjunto de datos
- Publisher: The Chinese University of Hong Kong et al.
- Size: 464510 instances
- License: CC-BY-4.0
- Source: Collected multi-lingual instructions, post-translated multi-lingual instructions, self-generated user-centered multi-lingual instructions
- Instruction Category: Multi
TigerBot_sft_en 2023-5 | Partial | EN | HG & CI & MC | Papel | Github | Conjunto de datos
- Publisher: TigerBot
- Size: 677117 instances
- License: Apache-2.0
- Source: Self-instruct, human-labeling, open-source data cleaning
- Instruction Category: Multi
TigerBot_sft_zh 2023-5 | Partial | ZH | HG & CI & MC | Papel | Github | Conjunto de datos
- Publisher: TigerBot
- Size: 530705 instances
- License: Apache-2.0
- Source: Self-instruct, human-labeling, open-source data cleaning
- Instruction Category: Multi
Aya Collection 2024-2 | All | Multi (114) | HG & CI & MC | Papel | Dataset | Sitio web
- Publisher: Cohere For AI Community et al.
- Size: 513M instances
- License: Apache-2.0
- Source: Templated data, Translated data and Aya Dataset
- Instruction Category: Multi
REInstruct 2024-8 | Not | EN | HG & CI & MC | Papel | Github
- Publisher: Chinese Information Processing Laboratory et al.
- Size: 35K instances
- License: -
- Source: Automatically constructing instruction data from the C4 corpus using a small amount of manually annotated seed instruction data
- Instruction Category: Multi

Domain-specific Instruction Fine-tuning Datasets

The domain-specific instruction fine-tuning datasets are constructed for a particular domain by formulating instructions that encapsulate knowledge and task types closely related to that domain.

Dataset information format：

 - Dataset name  Release Time | Public or Not | Language | Construction Method | Paper | Github | Dataset | Website
  - Publisher:
  - Size:
  - License:
  - Source:
  - Instruction Category:
  - Domain:

Médico

ChatDoctor 2023-3 | All | EN | HG & MC | Papel | Github | Conjunto de datos
- Publisher: University of Texas Southwestern Medical Center et al.
- Size: 115K instances
- License: Apache-2.0
- Source: Real conversations between doctors and patients & Generated by ChatGPT
- Instruction Category: Multi
- Domain: Medical
ChatMed_Consult_Dataset 2023-5 | All | ZH | MC | Github | Conjunto de datos
- Publisher: michael-wzhu
- Size: 549326 instances
- License: CC-BY-NC-4.0
- Source: Generated by GPT-3.5-Turbo
- Instruction Category: Multi
- Domain: Medical
CMtMedQA 2023-8 | All | ZH | HG | Papel | Github | Conjunto de datos
- Publisher: Zhengzhou University
- Size: 68023 instances
- License: MIT
- Source: Real conversations between doctors and patients
- Instruction Category: Multi
- Domain: Medical
DISC-Med-SFT 2023-8 | All | ZH | HG & CI | Papel | Github | Dataset | Sitio web
- Publisher: Fudan University et al.
- Size: 464898 instances
- License: Apache-2.0
- Source: Open source datasets & Manually selected data
- Instruction Category: Multi
- Domain: Medical
HuatuoGPT-sft-data-v1 2023-5 | All | ZH | HG & MC | Papel | Github | Conjunto de datos
- Publisher: The Chinese University of Hong Kong et al.
- Size: 226042 instances
- License: Apache-2.0
- Source: Real conversations between doctors and patients & Generated by ChatGPT
- Instruction Category: Multi
- Domain: Medical
Huatuo-26M 2023-5 | Partial | ZH | CI | Papel | Github
- Publisher: The Chinese University of Hong Kong et al.
- Size: 26504088 instances
- License: Apache-2.0
- Source: Collection and improvement of various datasets
- Instruction Category: Multi
- Domain: Medical
MedDialog 2020-4 | All | EN & ZH | HG | Papel | Github
- Publisher: UC San Diego
- Size: 3.66M instances
- License: -
- Source: Real conversations between doctors and patients
- Instruction Category: Multi
- Domain: Medical
Medical Meadow 2023-4 | All | EN | HG & CI | Papel | Github | Conjunto de datos
- Publisher: University Hospital Aachen et al.
- Size: 160076 instances
- License: GPL-3.0
- Source: Crawl data from the Internet & Collection and improvement of various NLP datasets
- Instruction Category: Multi
- Domain: Medical
Medical-sft 2023-5 | All | EN & ZH | CI | Github | Conjunto de datos
- Publisher: Ming Xu
- Size: 2.07M instances
- License: Apache-2.0
- Source: Collection and improvement of various NLP datasets
- Instruction Category: Multi
- Domain: Medical
QiZhenGPT-sft-20k 2023-5 | Partial | ZH | CI | Github | Conjunto de datos
- Publisher: Zhejiang University
- Size: 20K instances
- License: GPL-3.0
- Source: Collection and improvement of various datasets
- Instruction Category: Multi
- Domain: Medical
ShenNong_TCM_Dataset 2023-6 | All | ZH | MC | Github | Conjunto de datos
- Publisher: michael-wzhu
- Size: 112565 instances
- License: Apache-2.0
- Source: Generated by ChatGPT
- Instruction Category: Multi
- Domain: Medical

Código

Code_Alpaca_20K 2023-3 | All | EN & PL | MC | Github | Conjunto de datos
- Publisher: Sahil Chaudhary
- Size: 20K instances
- License: Apache-2.0
- Source: Generated by Text-Davinci-003
- Instruction Category: Code
- Domain: Code
CodeContest 2022-3 | All | EN & PL | CI | Papel | Github
- Publisher: DeepMind
- Size: 13610 instances
- License: Apache-2.0
- Source: Collection and improvement of various datasets
- Instruction Category: Code
- Domain: Code
CommitPackFT 2023-8 | All | EN & PL (277) | HG | Papel | Github | Conjunto de datos
- Publisher: Bigcode
- Size: 702062 instances
- License: MIT
- Source: GitHub Action dump
- Instruction Category: Code
- Domain: Code
ToolAlpaca 2023-6 | All | EN & PL | HG & MC | Papel | Github
- Publisher: Chinese Information Processing Laboratory et al.
- Size: 3928 instances
- License: Apache-2.0
- Source: Manually filter APIs & Generated by ChatGPT
- Instruction Category: Code
- Domain: Code
ToolBench 2023-7 | All | EN & PL | HG & MC | Papel | Github
- Publisher: Tsinghua University et al.
- Size: 126486 instances
- License: Apache-2.0
- Source: Manually filter APIs & Generated by ChatGPT
- Instruction Category: Code
- Domain: Code

Legal

DISC-Law-SFT 2023-9 | Partial | ZH | HG & CI & MC | Papel | Github | Sitio web
- Publisher: Fudan University et al.
- Size: 403K instances
- License: Apache-2.0
- Source: Open source datasets & Legal-related Text Content & Generated by GPT-3.5-Turbo
- Instruction Category: Multi
- Domain: Law
HanFei 1.0 2023-5 | All | ZH | - | Github | Conjunto de datos
- Publisher: Chinese Academy of Sciences et al.
- Size: 255K instances
- License: Apache-2.0
- Source: Filter legal-related data according to rules
- Instruction Category: Multi
- Domain: Law
LawGPT_zh 2023-5 | Partial | ZH | CI & MC | Github | Conjunto de datos
- Publisher: Shanghai Jiao Tong University
- Size: 200K instances
- License: -
- Source: Real conversations & Generated by ChatGPT
- Instruction Category: Multi
- Domain: Law
Lawyer LLaMA_sft 2023-5 | Partial | ZH | CI & MC | Papel | Github | Conjunto de datos
- Publisher: Peking Universit
- Size: 21476 instances
- License: Apache-2.0
- Source: Generated by ChatGPT with other datasets' prompts
- Instruction Category: Multi
- Domain: Law

Matemáticas

BELLE_School_Math 2023-5 | All | ZH | MC | Github | Conjunto de datos
- Publisher: BELLE
- Size: 248481 instances
- License: GPL-3.0
- Source: Generated by ChatGPT
- Instruction Category: Math
- Domain: Math
Goat 2023-5 | All | EN | HG | Papel | Github | Conjunto de datos
- Publisher: National University of Singapore
- Size: 1746300 instances
- License: Apache-2.0
- Source: Artificially synthesized data
- Instruction Category: Math
- Domain: Math
MWP 2021-9 | All | EN & ZH | CI | Papel | Github | Conjunto de datos
- Publisher: Xihua University et al.
- Size: 251598 instances
- License: MIT
- Source: Collection and improvement of various datasets
- Instruction Category: Math
- Domain: Math
OpenMathInstruct-1 2024-2 | All | EN | CI & MC | Papel | Github | Conjunto de datos
- Publisher: NVIDIA
- Size: 1.8M instances
- License: NVIDIA License
- Source: GSM8K and MATH datasets (original questions); Generated using Mixtral-8×7B model
- Instruction Category: Math
- Domain: Math

Educación

Child_chat_data 2023-8 | All | ZH | HG & MC | Github
- Publisher: Harbin Institute of Technology et al.
- Size: 5000 instances
- License: -
- Source: Real conversations & Generated by GPT-3.5-Turbo
- Instruction Category: Multi
- Domain: Education
Educhat-sft-002-data-osm 2023-7 | All | EN & ZH | CI | Papel | Github | Conjunto de datos
- Publisher: East China Normal University et al.
- Size: 4279419 instances
- License: CC-BY-NC-4.0
- Source: Collection and improvement of various datasets
- Instruction Category: Multi
- Domain: Education
TaoLi_data 2023-X | All | ZH | HG & CI | Github | Conjunto de datos
- Publisher: Beijing Language and Culture University et al.
- Size: 88080 instances
- License: Apache-2.0
- Source: Collection and improvement of various datasets & Manually extract dictionary data
- Instruction Category: Multi
- Domain: Education

Otro

DISC-Fin-SFT 2023-10 | Partial | ZH | HG & CI & MC | Papel | Github | Sitio web
- Publisher: Fudan University et al.
- Size: 246K instances
- License: Apache-2.0
- Source: Open source datasets & Manually collect financial data & ChatGPT assistance
- Instruction Category: Multi
- Domain: Financial
AlphaFin 2024-3 | All | EN & ZH | HG & CI & MC | Papel | Github | Conjunto de datos
- Publisher: South China University of Technology et al.
- Size: 167362 instances
- License: Apache-2.0
- Source: Traditional research datasets, real-time financial data, handwritten CoT data
- Instruction Category: Multi
- Domain: Financial
GeoSignal 2023-6 | Partial | EN | HG & CI & MC | Papel | Github | Conjunto de datos
- Publisher: Shanghai Jiao Tong University et al.
- Size: 22627272 instances
- License: Apache-2.0
- Source: Open source datasets & Geoscience-related Text Content & Generated by GPT-4
- Instruction Category: Multi
- Domain: Geoscience
MeChat 2023-4 | All | ZH | CI & MC | Papel | Github | Conjunto de datos
- Publisher: Zhejiang University et al.
- Size: 56K instances
- License: CC0-1.0
- Source: Based on PsyQA dataset with the proposed SMILE method
- Instruction Category: Multi
- Domain: Mental Health
Mol-Instructions 2023-6 | All | EN | HG & CI & MC | Papel | Github | Conjunto de datos
- Publisher: Zhejiang University et al.
- Size: 2043586 instances
- License: CC-BY-4.0
- Source: Molecule-oriented, Protein-oriented, Biomolecular text instructions
- Instruction Category: Multi
- Domain: Biology
Owl-Instruction 2023-9 | All | EN & ZH | HG & MC | Papel | Github
- Publisher: Beihang University et al.
- Size: 17858 instances
- License: -
- Source: Generated by GPT-4 & Manual verification
- Instruction Category: Multi
- Domain: IT
PROSOCIALDIALOG 2022-5 | All | EN | HG & MC | Papel | Conjunto de datos
- Publisher: Allenai
- Size: 165681 instances
- License: CC-BY-4.0
- Source: Generated by humans with GPT-3 created prompts
- Instruction Category: Social Norms
- Domain: Social Norms
TransGPT-sft 2023-7 | All | ZH | HG | Github | Conjunto de datos
- Publisher: Beijing Jiaotong University
- Size: 58057 instances
- License: Apache-2.0
- Source: Manually collect traffic-related data
- Instruction Category: Multi
- Domain: Transportation

Preference Datasets

Preference datasets are collections of instructions that provide preference evaluations for multiple responses to the same instruction input.

Preference Evaluation Methods

The preference evaluation methods for preference datasets can be categorized into voting, sorting, scoring, and other methods. Datasets are classified based on preference evaluation methods.

Dataset information format：

 - Dataset name  Release Time | Public or Not | Language | Construction Method | Paper | Github | Dataset | Website
  - Publisher:
  - Size:
  - License:
  - Domain:
  - Instruction Category: 
  - Preference Evaluation Method: 
  - Source:

Votar

Chatbot_arena_conversations 2023-6 | All | Multi | HG & MC | Papel | Conjunto de datos
- Publisher: UC Berkeley et al.
- Size: 33000 instances
- License: CC-BY-4.0 & CC-BY-NC-4.0
- Domain: General
- Instruction Category: Multi
- Preference Evaluation Method: VO-H
- Source: Generated by twenty LLMs & Manual judgment
hh-rlhf 2022-4 | All | EN | HG & MC | Paper1 | Paper2 | Github | Conjunto de datos
- Publisher: Anthropic
- Size: 169352 instances
- License: MIT
- Domain: General
- Instruction Category: Multi
- Preference Evaluation Method: VO-H
- Source: Generated by LLMs & Manual judgment
MT-Bench_human_judgments 2023-6 | All | EN | HG & MC | Papel | Github | Dataset | Sitio web
- Publisher: UC Berkeley et al.
- Size: 3.3K instances
- License: CC-BY-4.0
- Domain: General
- Instruction Category: Multi
- Preference Evaluation Method: VO-H
- Source: Generated by LLMs & Manual judgment
PKU-SafeRLHF 2023-7 | Partial | EN | HG & CI & MC | Papel | Github | Conjunto de datos
- Publisher: Peking University
- Size: 361903 instances
- License: CC-BY-NC-4.0
- Domain: Social Norms
- Instruction Category: Social Norms
- Preference Evaluation Method: VO-H
- Source: Generated by LLMs & Manual judgment
SHP 2021-10 | All | EN | HG | Papel | Github | Conjunto de datos
- Publisher: Stanford
- Size: 385563 instances
- License: -
- Domain: General
- Instruction Category: Multi
- Preference Evaluation Method: VO-H
- Source: Reddit data & Manual judgment
Zhihu_rlhf_3k 2023-4 | All | ZH | HG | Conjunto de datos
- Publisher: Liyucheng
- Size: 3460 instances
- License: CC-BY-2.0
- Domain: General
- Instruction Category: Multi
- Preference Evaluation Method: VO-H
- Source: Zhihu data & Manual judgment
Summarize_from_Feedback 2020-9 | All | EN | HG & CI | Papel | Conjunto de datos
- Publisher: OpenAI
- Size: 193841 instances
- License: -
- Domain: News
- Instruction Category: Multi
- Preference Evaluation Method: VO-H & SC-H
- Source: Open source datasets & Manual judgment and scoring
CValues 2023-7 | All | ZH | MC | Papel | Github | Conjunto de datos
- Publisher: Alibaba Group
- Size: 145K instances
- License: Apache-2.0
- Domain: Social Norms
- Instruction Category: Social Norms
- Preference Evaluation Method: VO-M
- Source: Generated by LLMs & Evaluation by the reward model
huozi_rlhf_data 2024-2 | All | ZH | HG & MC | Github | Conjunto de datos
- Publisher: Huozi-Team
- Size: 16918 instances
- License: Apache-2.0
- Domain: General
- Instruction Category: Multi
- Preference Evaluation Method: VO-H
- Source: Generated by Huozi model & Manual judgment

Clasificar

OASST1_pairwise_rlhf_reward 2023-5 | All | Multi | CI | Conjunto de datos
- Publisher: Tasksource
- Size: 18918 instances
- License: Apache-2.0
- Domain: General
- Instruction Category: Multi
- Preference Evaluation Method: SO-H
- Source: OASST1 datasets & Manual sorting

Puntaje

Stack-Exchange-Preferences 2021-12 | All | EN | HG | Papel | Conjunto de datos
- Publisher: Anthropic
- Size: 10807695 instances
- License: CC-BY-SA-4.0
- Domain: General
- Instruction Category: Multi
- Preference Evaluation Method: SC-H
- Source: Stackexchange data & Manual scoring
WebGPT 2021-12 | All | EN | HG & CI | Papel | Conjunto de datos
- Publisher: OpenAI
- Size: 19578 instances
- License: -
- Domain: General
- Instruction Category: Multi
- Preference Evaluation Method: SC-H
- Source: Open source datasets & Manual scoring
Alpaca_comparison_data 2023-3 | All | EN | MC | Github
- Publisher: Stanford Alpaca
- Size: 51K instances
- License: Apache-2.0
- Domain: General
- Instruction Category: Multi
- Preference Evaluation Method: SC-M
- Source: Generated by three LLMs & GPT-4 scoring
Stable_Alignment 2023-5 | All | EN | MC | Papel | Github
- Publisher: Dartmouth College et al.
- Size: 169K instances
- License: Apache-2.0
- Domain: General
- Instruction Category: Multi
- Preference Evaluation Method: SC-M
- Source: Generated by LLMs & Model scoring
UltraFeedback 2023-10 | All | EN | CI & MC | Papel | Github | Conjunto de datos
- Publisher: Tsinghua University et al.
- Size: 63967 instances
- License: MIT
- Domain: General
- Instruction Category: Multi
- Preference Evaluation Method: SC-M
- Source: Generated by seventeen LLMs & Model scoring
OpenHermesPreferences 2024-2 | All | EN | CI & MC | Conjunto de datos
- Publisher: Argilla et al.
- Size: 989490 instances
- License: -
- Domain: General
- Instruction Category: Multi
- Preference Evaluation Method: SO-M
- Source: OpenHermes-2.5 dataset & Model sorting
HelpSteer 2023-11 | All | EN | HG & CI & MC | Papel | Conjunto de datos
- Publisher: NVIDIA
- Size: 37120 instances
- License: CC-BY-4.0
- Domain: General
- Instruction Category: Multi
- Preference Evaluation Method: SC-H
- Source: Generated by LLMs & Manual judgment
HelpSteer2 2024-6 | All | EN | HG & CI & MC | Papel | Github | Conjunto de datos
- Publisher: NVIDIA
- Size: 21362 instances
- License: CC-BY-4.0
- Domain: General
- Instruction Category: Multi
- Preference Evaluation Method: SC-H
- Source: Generated by LLMs & Manual judgment

Otro

Medical-rlhf 2023-5 | All | ZH | CI & MC | Github | Conjunto de datos
- Publisher: Ming Xu
- Size: 4K instances
- License: Apache-2.0
- Domain: Medical
- Instruction Category: Multi
- Preference Evaluation Method: Other
- Source: Response_chosen comes from the doctor's response & Response_rejected comes from the model's response
PRM800K 2023-5 | All | EN | HG & CI & MC | Papel | Github
- Publisher: OpenAI
- Size: 800K instances
- License: MIT
- Domain: Math
- Instruction Category: Math
- Preference Evaluation Method: Other
- Source: Generated by LLMs & Mathematical reasoning steps are determined manually

Evaluation Datasets

Evaluation datasets are a carefully curated and annotated set of data samples used to assess the performance of LLMs across various tasks. Datasets are classified based on evaluation domains.

Dataset information format：

 - Dataset name  Release Time | Public or Not | Language | Construction Method | Paper | Github | Dataset | Website
  - Publisher:
  - Size:
  - License:
  - Question Type: 
  - Evaluation Method: 
  - Focus: 
  - Numbers of Evaluation Categories/Subcategories: 
  - Evaluation Category:

General

AlpacaEval 2023-5 | All | EN | CI & MC | Papel | Github | Dataset | Sitio web
- Publisher: Stanford et al.
- Size: 805 instances
- License: Apache-2.0
- Question Type: SQ
- Evaluation Method: ME
- Focus: The performance on open-ended question answering
- Numbers of Evaluation Categories/Subcategories: 1/-
- Evaluation Category: Open-ended question answering
BayLing-80 2023-6 | All | EN & ZH | HG & CI | Papel | Github | Conjunto de datos
- Publisher: Chinese Academy of Sciences
- Size: 320 instances
- License: GPL-3.0
- Question Type: SQ
- Evaluation Method: ME
- Focus: Chinese-English language proficiency and multimodal interaction skills
- Numbers of Evaluation Categories/Subcategories: 9/-
- Evaluation Category: Writing, Roleplay, Common-sense, Fermi, Counterfactual, Coding, Math, Generic, Knowledge
BELLE_eval 2023-4 | All | ZH | HG & MC | Papel | Github
- Publisher: BELLE
- Size: 1000 instances
- License: Apache-2.0
- Question Type: SQ
- Evaluation Method: ME
- Focus: The performance of Chinese language models in following instructions
- Numbers of Evaluation Categories/Subcategories: 9/-
- Evaluation Category: Extract, Closed qa, Rewrite, Summarization, Generation, Classification, Brainstorming, Open qa, Others
CELLO 2023-9 | All | EN | HG | Papel | Github
- Publisher: Fudan University et al.
- Size: 523 instances
- License: -
- Question Type: SQ
- Evaluation Method: CE
- Focus: The ability of LLMs to understand complex instructions
- Numbers of Evaluation Categories/Subcategories: 2/10
- Evaluation Category: Complex task description, Complex input
MT-Bench 2023-6 | All | EN | HG | Papel | Github | Sitio web
- Publisher: UC Berkeley et al.
- Size: 80 instances
- License: Apache-2.0
- Question Type: SQ
- Evaluation Method: ME
- Focus: The performance on open-ended question answering
- Numbers of Evaluation Categories/Subcategories: 8/-
- Evaluation Category: Writing, Roleplay, Reasoning, Math, Coding, Extraction, STEM, Humanities
SuperCLUE 2023-7 | Not | ZH | HG & MC | Papel | Github | Website1 | Website2
- Publisher: CLUE et al.
- Size: 3754 instances
- License: -
- Question Type: Multi
- Evaluation Method: HE & CE
- Focus: The performance in a Chinese context
- Numbers of Evaluation Categories/Subcategories: 2/-
- Evaluation Category: Open multi-turn open questions, OPT objective questions
Vicuna Evaluation 2023-3 | All | EN | HG | Github | Dataset | Sitio web
- Publisher: LMSYS ORG
- Size: 80 instances
- License: Apache-2.0
- Question Type: SQ
- Evaluation Method: ME
- Focus: The performance on open-ended question answering
- Numbers of Evaluation Categories/Subcategories: 9/-
- Evaluation Category: Generic, Knowledge, Roleplay, Common-sense, Fermi, Counterfactual, Coding, Math, Writing
CHC-Bench 2024-4 | All | ZH | HG & CI | Papel | Github | Dataset | Sitio web
- Publisher: Multimodal Art Projection Research Community et al.
- Size: 214 instances
- License: Apache-2.0
- Question Type: Multi
- Evaluation Method: ME
- Focus: Hard-case Chinese instructions understanding and following
- Numbers of Evaluation Categories/Subcategories: 8/-
- Evaluation Category: Writing, Humanity, Science, Role-playing, Reading Comprehension, Math, Hard Cases, Coding
CIF-Bench 2024-2 | Partial | ZH | HG & CI | Papel | Github | Sitio web
- Publisher: University of Manchester et al.
- Size: 15K instances
- License: -
- Question Type: SQ
- Evaluation Method: CE & ME
- Focus: Evaluate the zero-shot generalizability of LLMs to the Chinese language
- Numbers of Evaluation Categories/Subcategories: 10/150
- Evaluation Category: Chinese culture, Classification, Code, Commonsense, Creative NLG, Evaluation, Grammar, Linguistic, Motion detection, NER
WildBench 2024-6 | All | EN | HG & CI | Papel | Github | Dataset | Sitio web
- Publisher: Allen Institute for AI et al.
- Size: 1024 instances
- License: AI2 ImpACT License
- Question Type: SQ
- Evaluation Method: ME
- Focus: An automated evaluation framework designed to benchmark LLMs using challenging, real-world user queries.
- Numbers of Evaluation Categories/Subcategories: 11/-
- Evaluation Category: Information seeking, Coding & Debugging, Creative writing, Reasoning, Planning, Math, Editing, Data analysis, Role playing, Brainstorming, Advice seeking
SysBench 2024-8 | All | EN | HG | Papel | Github | Conjunto de datos
- Publisher: Peking University et al.
- Size: 500 instances
- License: -
- Question Type: SQ
- Evaluation Method: ME
- Focus: Systematically analyze system message following ability
- Numbers of Evaluation Categories/Subcategories: 3/-
- Evaluation Category: Constraint complexity, Instruction misalignment, Multi-turn stability

Examen

AGIEval 2023-4 | All | EN & ZH | HG & CI | Papel | Github | Conjunto de datos
- Publisher: Microsoft
- Size: 8062 instances
- License: MIT
- Question Type: OQ
- Evaluation Method: CE
- Focus: Human-centric standardized exams
- Numbers of Evaluation Categories/Subcategories: 7/20
- Evaluation Category: Gaokao, SAT, JEC, LSAT, LogiQA, AQuA-RAT, Math
GAOKAO-Bench 2023-5 | All | ZH | HG | Papel | Github
- Publisher: Fudan University et al.
- Size: 2811 instances
- License: Apache-2.0
- Question Type: Multi
- Evaluation Method: HE & CE
- Focus: Chinese Gaokao examination
- Numbers of Evaluation Categories/Subcategories: 10/-
- Evaluation Category: Chinese, Mathematics (2 categories), English, Physics, Chemistry, Biology, Politics, History, Geography
M3Exam 2023-6 | All | Multi (9) | HG | Paper | Github
- Publisher: Alibaba Group et al.
- Size: 12317 instances
- License: -
- Question Type: OQ
- Evaluation Method: CE
- Focus: The comprehensive abilities in a multilingual and multilevel context using real human exam questions
- Numbers of Evaluation Categories/Subcategories: 3/-
- Evaluation Category: Low, Mid, High

Sujeto

ARB 2023-7 | All | EN | CI | Paper | Github
- Publisher: DuckAI et al.
- Size: 1207 instances
- License: MIT
- Question Type: Multi
- Evaluation Method: HE & ME
- Focus: Advanced reasoning problems in multiple fields
- Numbers of Evaluation Categories/Subcategories: 5/-
- Evaluation Category: Mathematics, Physics, Law, MCAT(Reading), MCAT(Science)
C-CLUE 2021-8 | All | ZH | HG | Github | Sitio web
- Publisher: Tianjin University
- Tamaño: -
- License: CC-BY-SA-4.0
- Question Type: SQ
- Evaluation Method: CE
- Focus: Classical Chinese language understanding
- Numbers of Evaluation Categories/Subcategories: 2/-
- Evaluation Category: Named entity recognition, Relation extraction
C-Eval 2023-5 | All | ZH | HG & MC | Paper | Github | Dataset | Sitio web
- Publisher: Shanghai Jiao Tong University
- Size: 13948 instances
- License: CC-BY-NC-SA-4.0
- Question Type: OQ
- Evaluation Method: CE
- Focus: The advanced knowledge and reasoning abilities in a Chinese context
- Numbers of Evaluation Categories/Subcategories: 4/52
- Evaluation Category: STEM, Social Science, Humanity, Other
CG-Eval 2023-8 | All | ZH | HG | Paper | Github | Dataset | Sitio web
- Publisher: LanguageX AI Lab et al.
- Size: 11000 instances
- License: CC-BY-SA-4.0
- Question Type: SQ
- Evaluation Method: CE
- Focus: The generation capabilities of LLMs across various academic disciplines
- Numbers of Evaluation Categories/Subcategories: 6/55
- Evaluation Category: Science and engineering, Humanities and social sciences, Mathematical calculations, Medical practitioner qualification Examination, Judicial Examination, Certfied public accountant examination
LLMEVAL-3 2023-9 | Not | ZH | HG | Github | Sitio web
- Publisher: Fudan University et al.
- Size: 200K instances
- License: -
- Question Type: SQ
- Evaluation Method: ME
- Focus: Subject-specific knowledge capability
- Numbers of Evaluation Categories/Subcategories: 13/-
- Evaluation Category: Philosophy, Economics, Law, Education, Literature, History, Science, Engineering, Agriculture, Medicine, Military science, Management, Fine arts
MMCU 2023-4 | All | ZH | HG | Paper | Github
- Publisher: LanguageX AI Lab
- Size: 11845 instances
- License: -
- Question Type: OQ
- Evaluation Method: CE
- Focus: Multidisciplinary abilities
- Numbers of Evaluation Categories/Subcategories: 4/25
- Evaluation Category: Medicine, Law, Psychology, Education
MMLU 2020-9 | All | EN | HG | Paper | Github
- Publisher: UC Berkeley et al.
- Size: 15908 instances
- License: MIT
- Question Type: OQ
- Evaluation Method: CE
- Focus: Knowledge in academic and professional domains
- Numbers of Evaluation Categories/Subcategories: 4/57
- Evaluation Category: Humanities, Social science, STEM, Other
M3KE 2023-5 | All | ZH | HG | Papel | Github | Conjunto de datos
- Publisher: Tianjin University et al.
- Size: 20477 instances
- License: Apache-2.0
- Question Type: OQ
- Evaluation Method: CE
- Focus: Multidisciplinary abilities
- Numbers of Evaluation Categories/Subcategories: 4/71
- Evaluation Category: Arts & Humanities, Social sciences, Natural sciences, Other
SCIBENCH 2023-7 | All | EN | HG | Paper | Github
- Publisher: University of California et al.
- Size: 695 instances
- License: MIT
- Question Type: SQ
- Evaluation Method: CE
- Focus: The performance in university-level science and engineering domains
- Numbers of Evaluation Categories/Subcategories: 3/10
- Evaluation Category: Physics, Chemistry, Math
ScienceQA 2022-9 | All | EN | HG | Paper | Github | Sitio web
- Publisher: University of California et al.
- Size: 21208 instances
- License: CC-BY-NC-SA-4.0
- Question Type: OQ
- Evaluation Method: CE
- Focus: Science question-answering ability
- Numbers of Evaluation Categories/Subcategories: 3/26
- Evaluation Category: Natural science, Social science, Language science
TheoremQA 2023-5 | All | EN | HG | Paper | Github | Conjunto de datos
- Publisher: University of Waterloo et al.
- Size: 800 instances
- License: MIT
- Question Type: SQ
- Evaluation Method: CE
- Focus: Science subject question-answering ability
- Numbers of Evaluation Categories/Subcategories: 4/39
- Evaluation Category: Mathematics, Physics, Finance, CS & EE
XiezhiBenchmark 2023-6 | All | EN & ZH | HG & MC | Papel | Github
- Publisher: Fudan University et al.
- Size: 249587 instances
- License: CC-BY-NC-SA-4.0
- Question Type: OQ
- Evaluation Method: CE
- Focus: Multidisciplinary abilities
- Numbers of Evaluation Categories/Subcategories: 13/516
- Evaluation Category: Medicine, Literature, Economics, Agronomy, Science, Jurisprudence, History, Art studies, Philosophy, Pedagogy, Military science, Management, Engineering
CMMLU 2023-6 | All | ZH | HG | Paper | Github | Conjunto de datos
- Publisher: MBZUAI
- Size: 11528 instances
- License: CC-BY-NC-4.0
- Question Type: OQ
- Evaluation Method: CE
- Focus: The knowledge and reasoning capabilities within the Chinese context
- Numbers of Evaluation Categories/Subcategories: 5/67
- Evaluation Category: Social science, STEM, Humanities, China specific, Other
GPQA 2023-11 | All | EN | HG | Paper | Github | Conjunto de datos
- Publisher: New York University et al.
- Size: 448 instances
- License: CC-BY-4.0
- Question Type: OQ
- Evaluation Method: CE
- Focus: The disciplinary knowledge in the fields of biology, physics, and chemistry
- Numbers of Evaluation Categories/Subcategories: 3/16
- Evaluation Category: Biology, Physics, Chemistry
CMATH 2023-6 | All | ZH | HG | Paper | Github | Conjunto de datos
- Publisher: Xiaomi AI Lab
- Size: 1698 instances
- License: CC-BY-4.0
- Question Type: SQ
- Evaluation Method: CE
- Focus: Elementary school math word problems
- Numbers of Evaluation Categories/Subcategories: 6/-
- Evaluation Category: Grades 1 to 6 in elementary school
FineMath 2024-3 | Not | ZH | HG | Papel
- Publisher: Tianjin University et al.
- Size: 1584 instances
- License: -
- Question Type: Multi
- Evaluation Method: -
- Focus: Elementary school math word problems
- Numbers of Evaluation Categories/Subcategories: 6/17
- Evaluation Category: Number & Operations, Measurement, Data analysis & Probability, Algebra, Geometry, Others
WYWEB 2023-7 | All | ZH | HG & CI | Paper | Github | Dataset | Sitio web
- Publisher: Zhejiang University et al.
- Size: 467200 instances
- License: -
- Question Type: Multi
- Evaluation Method: CE
- Focus: Classical Chinese
- Numbers of Evaluation Categories/Subcategories: 5/9
- Evaluation Category: Sequence labeling, Sentence classification, Token similarity, Reading comprehension, Translation
ACLUE 2023-10 | All | ZH | HG & CI | Paper | Github | Conjunto de datos
- Publisher: Mohamed bin Zayed University of Artificial Intelligence
- Size: 4967 instances
- License: CC-BY-NC-SA-4.0
- Question Type: OQ
- Evaluation Method: CE
- Focus: Classical Chinese language understanding
- Numbers of Evaluation Categories/Subcategories: 5/15
- Evaluation Category: Lexical, Syntactic, Semantic, Inference, Knowledge
SciKnowEval 2024-6 | All | EN | HG & CI & MC | Papel | Github | Conjunto de datos
- Publisher: Zhejiang University et al.
- Size: 50048 instances
- License: -
- Question Type: Multi
- Evaluation Method: CE & ME
- Focus: Evaluate the capabilities of LLMs in handling scientific knowledge
- Numbers of Evaluation Categories/Subcategories: 2/49
- Evaluation Category: Biology, Chemistry
C ³ Bench 2024-5 | All | ZH | HG & CI | Papel
- Publisher: South China University of Technology
- Size: 50000 instances
- License: -
- Question Type: SQ
- Evaluation Method: CE
- Focus: Classical Chinese
- Numbers of Evaluation Categories/Subcategories: 5/-
- Evaluation Category: Classification, Retrieval, NER, Punctuation, Translation
ArabicMMLU 2024-8 | All | AR | HG | Paper | Github | Conjunto de datos
- Publisher: MBZUAI et al.
- Size: 14575 instances
- License: CC-BY-NC-SA-4.0
- Question Type: OQ
- Evaluation Method: CE
- Focus: Multi-task language understanding benchmark for the Arabic language
- Numbers of Evaluation Categories/Subcategories: 5/40
- Evaluation Category: STEM, Social science, Humanities, Language, Other
PersianMMLU 2024-4 | All | FA | HG | Paper | Conjunto de datos
- Publisher: Raia Center for Artificial Intelligence Research et al.
- Size: 20192 instances
- License: CC-ND
- Question Type: OQ
- Evaluation Method: CE
- Focus: Facilitate the rigorous evaluation of LLMs that support the Persian language
- Numbers of Evaluation Categories/Subcategories: 5/38
- Evaluation Category: Social science, Humanities, Natural science, Mathematics, Other
TMMLU+ 2024-3 | All | ZH | HG & CI | Paper | Conjunto de datos
- Publisher: iKala AI Lab et al.
- Size: 22690 instances
- License: MIT
- Question Type: OQ
- Evaluation Method: CE
- Focus: Evaluate the language understanding capabilities in Traditional Chinese
- Numbers of Evaluation Categories/Subcategories: 4/66
- Evaluation Category: STEM, Social sciences, Humanities, Other

NLU

CLUE 2020-12 | All | ZH | CI | Paper | Github
- Publisher: CLUE team
- Size: 9 datasets
- License: -
- Question Type: SQ
- Evaluation Method: CE
- Focus: Natural language understanding capability
- Numbers of Evaluation Categories/Subcategories: 3/9
- Evaluation Category: Single-sentence tasks, Sentence pair tasks, Machine reading comprehension tasks
CUGE 2021-12 | All | EN & ZH | CI | Paper | Sitio web
- Publisher: Tsinghua University et al.
- Size: 33.4M instances
- License: -
- Question Type: SQ
- Evaluation Method: CE
- Focus: Natural language understanding capability
- Numbers of Evaluation Categories/Subcategories: 7/18
- Evaluation Category: Language understanding (word-sentence or discourse level), Information acquisition and question answering, Language generation, Conversational interaction, Multilingual, Mathematical reasoning
GLUE 2018-11 | All | EN | CI | Paper | Github | Sitio web
- Publisher: New York University et al.
- Size: 9 datasets
- License: -
- Question Type: SQ
- Evaluation Method: CE
- Focus: Natural language understanding capability
- Numbers of Evaluation Categories/Subcategories: 3/9
- Evaluation Category: Single-sentence tasks, Similarity and paraphrase tasks, Inference tasks
SuperGLUE 2019-5 | All | EN | CI | Paper | Sitio web
- Publisher: New York University et al.
- Size: 8 datasets
- License: -
- Question Type: SQ
- Evaluation Method: CE
- Focus: Natural language understanding capability
- Numbers of Evaluation Categories/Subcategories: 4/8
- Evaluation Category: Word sense disambiguation, Natural language inference, Coreference resolution, Question answering
MCTS 2023-6 | All | ZH | HG | Paper | Github
- Publisher: Beijing Language and Culture University
- Size: 723 instances
- License: -
- Question Type: SQ
- Evaluation Method: CE
- Focus: Text simplification ability
- Numbers of Evaluation Categories/Subcategories: 1/-
- Evaluation Category: Text simplification
RAFT 2021-9 | All | EN | HG & CI | Paper | Dataset | Sitio web
- Publisher: Ought et al.
- Size: 28712 instances
- License: -
- Question Type: SQ
- Evaluation Method: CE
- Focus: Text classification ability
- Numbers of Evaluation Categories/Subcategories: 1/11
- Evaluation Category: Text classification
SentEval 2018-5 | All | EN | CI | Paper | Github
- Publisher: Facebook Artificial Intelligence Research
- Size: 28 datasets
- License: BSD
- Question Type: SQ
- Evaluation Method: CE
- Focus: The quality of universal sentence representations
- Numbers of Evaluation Categories/Subcategories: 1/21
- Evaluation Category: Universal sentence representations
LeSC 2024-5 | All | EN & ZH | HG | Paper | Github | Conjunto de datos
- Publisher: Tsinghua University et al.
- Size: 600 instances
- License: MIT
- Question Type: OQ
- Evaluation Method: CE
- Focus: The genuine linguistic-cognitive skills of LLMs
- Numbers of Evaluation Categories/Subcategories: 1/-
- Evaluation Category: Polysemy
KoBEST 2022-10 | All | KO | CI | Paper | Conjunto de datos
- Publisher: University of Oxford et al.
- Size: 5 datasets
- License: CC-BY-SA-4.0
- Question Type: OQ
- Evaluation Method: CE
- Focus: Korean balanced evaluation of significant tasks
- Numbers of Evaluation Categories/Subcategories: 5/-
- Evaluation Category: KB-BoolQ, KB-COPA, KB-WiC, KB-HellaSwag, KB-SentiNeg
SarcasmBench 2024-8 | All | EN | CI | Papel
- Publisher: Tianjin University et al.
- Size: 58347 instances
- License: -
- Question Type: SQ
- Evaluation Method: CE
- Focus: Evaluate LLMs on sarcasm understanding
- Numbers of Evaluation Categories/Subcategories: 1/-
- Evaluation Category: Sarcasm understanding

Razonamiento

Chain-of-Thought Hub 2023-5 | All | EN | CI | Papel | Github
- Publisher: University of Edinburgh et al.
- Tamaño: -
- License: MIT
- Question Type: SQ
- Evaluation Method: CE
- Focus: The multi-step reasoning capabilities
- Numbers of Evaluation Categories/Subcategories: 6/8
- Evaluation Category: Math, Science, Symbolic, Knowledge, Coding, Factual
Choice-75 2023-9 | All | EN | HG & CI & MC | Paper | Github
- Publisher: University of Pittsburgh et al.
- Size: 650 instances
- License: -
- Question Type: OQ
- Evaluation Method: CE
- Focus: Predict decisions based on descriptive scenarios
- Numbers of Evaluation Categories/Subcategories: 4/-
- Evaluation Category: Easy, Medium, Hard, N/A
NeuLR 2023-6 | All | EN | CI | Paper | Github | Conjunto de datos
- Publisher: Xi'an Jiaotong University et al.
- Size: 3000 instances
- License: -
- Question Type: SQ
- Evaluation Method: CE
- Focus: Logical reasoning capabilities
- Numbers of Evaluation Categories/Subcategories: 3/-
- Evaluation Category: Deductive, Inductive, Abductive
TabMWP 2022-9 | All | EN | HG | Paper | Github | Sitio web
- Publisher: University of California et al.
- Size: 38431 instances
- License: CC-BY-NC-SA-4.0
- Question Type: Multi
- Evaluation Method: CE
- Focus: Mathematical reasoning ability involving both textual and tabular information
- Numbers of Evaluation Categories/Subcategories: 1/-
- Evaluation Category: Mathematical reasoning and table QA
LILA 2022-10 | All | EN | CI | Paper | Github | Conjunto de datos
- Publisher: Arizona State Univeristy et al.
- Size: 317262 instances
- License: CC-BY-4.0
- Question Type: Multi
- Evaluation Method: CE
- Focus: Mathematical reasoning across diverse tasks
- Numbers of Evaluation Categories/Subcategories: 4/23
- Evaluation Category: Math ability, Language, Knowledge, Format
MiniF2F_v1 2021-9 | All | EN | HG & CI | Paper | Github
- Publisher: Ecole Polytechnique et al.
- Size: 488 instances
- License: -
- Question Type: SQ
- Evaluation Method: CE
- Focus: The performance on formal Olympiad-level mathematics problem statements
- Numbers of Evaluation Categories/Subcategories: 1/-
- Evaluation Category: Math
GameBench 2024-6 | All | EN | HG | Paper | Github | Conjunto de datos
- Publisher: Olin College of Engineering et al.
- Size: 9 Games
- License: CC-BY
- Question Type: SQ
- Evaluation Method: CE
- Focus: Evaluate strategic reasoning abilities of LLM agents
- Numbers of Evaluation Categories/Subcategories: 6/9
- Evaluation Category: Abstract Strategy, Non-Deterministic, Hidden Information, Language Communication, Social Deduction, Cooperation
TableBench 2024-8 | All | EN | HG & CI & MC | Paper | Github | Dataset | Sitio web
- Publisher: Beihang University et al.
- Size: 886 instances
- License: Apache-2.0
- Question Type: SQ
- Evaluation Method: CE
- Focus: Table question answering (TableQA) capabilities
- Numbers of Evaluation Categories/Subcategories: 4/18
- Evaluation Category: Fact checking, Numerical reasoning, Data analysis, Visualization

Conocimiento

ALCUNA 2023-10 | All | EN | HG | Paper | Github | Conjunto de datos
- Publisher: Peking University
- Size: 84351 instances
- License: MIT
- Question Type: Multi
- Evaluation Method: CE
- Focus: Assess the ability of LLMs to respond to new knowledge
- Numbers of Evaluation Categories/Subcategories: 3/-
- Evaluation Category: Knowledge understanding, Knowledge differentiation, Knowledge association
KoLA 2023-6 | Partial | EN | HG & CI | Paper | Github | Sitio web
- Publisher: Tsinghua University
- Size: 2138 instances
- License: GPL-3.0
- Question Type: SQ
- Evaluation Method: CE
- Focus: The ability to grasp and utilize world knowledge
- Numbers of Evaluation Categories/Subcategories: 4/19
- Evaluation Category: Knowledge memorization, Knowledge understanding, Knowledge applying, Knowledge creating
LLMEVAL-2 2023-7 | All | ZH | HG | Github
- Publisher: Fudan University et al.
- Size: 480 instances
- License: -
- Question Type: Multi
- Evaluation Method: HE & ME
- Focus: Knowledge capability
- Numbers of Evaluation Categories/Subcategories: 12/-
- Evaluation Category: Computer science, Economics, Foreign languages, Law, Mathematics, Medicine, Optics, Physics, Social sciences, Chinese language and literature, Chemistry, Life sciences
SocKET 2023-5 | All | EN | CI | Paper | Github
- Publisher: University of Michigan et al.
- Size: 2616342 instances
- License: CC-BY-4.0
- Question Type: SQ
- Evaluation Method: CE
- Focus: Mastery of social knowledge
- Numbers of Evaluation Categories/Subcategories: 4/58
- Evaluation Category: Classification, Regression, Pair-wise comparison, Span identification
LMExamQA 2023-6 | All | EN | MC | Paper | Sitio web
- Publisher: Tsinghua University et al.
- Size: 10090 instances
- License: -
- Question Type: SQ
- Evaluation Method: ME
- Focus: The performance on open-ended question answering
- Numbers of Evaluation Categories/Subcategories: 3/25
- Evaluation Category: Knowledge memorization, Knowledge comprehension, Knowledge analysis
DebateQA 2024-8 | All | EN | HG & CI & MC | Papel | Github | Conjunto de datos
- Publisher: Tsinghua Universty et al.
- Size: 2941 instances
- License: -
- Question Type: SQ
- Evaluation Method: ME
- Focus: Evaluate the comprehensiveness of perspectives and assess whether the LLM acknowledges the question's debatable nature
- Numbers of Evaluation Categories/Subcategories: 2/-
- Evaluation Category: Perspective diversity, Dispute awareness

Long Text

L-Eval 2023-7 | All | EN | HG & CI | Paper | Github | Dataset
- Publisher: Fudan University et al.
- Size: 2043 instances
- License: GPL-3.0
- Question Type: SQ
- Evaluation Method: HE & CE & ME
- Focus: Long text task capability
- Numbers of Evaluation Categories/Subcategories: 1/18
- Evaluation Category: Long text task
LongBench 2023-8 | All | EN & ZH | CI | Paper | Github | Conjunto de datos
- Publisher: Tsinghua University et al.
- Size: 4750 instances
- License: MIT
- Question Type: SQ
- Evaluation Method: CE
- Focus: Long text task capability
- Numbers of Evaluation Categories/Subcategories: 6/21
- Evaluation Category: Single-doc QA, Multi-doc QA, Summarization, Few-shot learning, Synthetic tasks, Code completion
LongEval 2023-6 | All | EN | HG | Github | Sitio web
- Publisher: LMSYS
- Tamaño: -
- License: Apache-2.0
- Question Type: SQ
- Evaluation Method: CE
- Focus: Long text task capability
- Numbers of Evaluation Categories/Subcategories: 2/-
- Evaluation Category: Coarse-grained topic retrieval, Fine-grained line retrieval
InfiniteBench 2023-11 | All | EN & ZH | HG & CI & MC | Github | Dataset
- Publisher: Tsinghua University et al.
- Size: 3932 instances
- License: Apache-2.0
- Question Type: Multi
- Evaluation Method: -
- Focus: Long text task capability
- Numbers of Evaluation Categories/Subcategories: 5/12
- Evaluation Category: Mathematics, Code, Dialogue, Books, Retrieval
ZeroSCROLLS 2023-5 | All | EN | HG & CI | Paper | Github | Dataset | Sitio web
- Publisher: Tel Aviv University et al.
- Size: 4378 instances
- License: MIT
- Question Type: Multi
- Evaluation Method: CE
- Focus: Long text task capability
- Numbers of Evaluation Categories/Subcategories: 3/10
- Evaluation Category: Summarization, Question Answering, Aggregation
LooGLE 2023-11 | All | EN | HG & CI & MC | Paper | Github | Dataset
- Publisher: BIGAI et al.
- Size: 6448 instances
- License: CC-BY-SA-4.0
- Question Type: SQ
- Evaluation Method: HE & CE & ME
- Focus: Long text task capability
- Numbers of Evaluation Categories/Subcategories: 2/4
- Evaluation Category: Long dependency tasks, Short dependency tasks
NAH (Needle-in-a-Haystack) 2023-11 | All | EN | - | Github
- Publisher: gkamradt et al.
- Tamaño: -
- License: MIT
- Question Type: SQ
- Evaluation Method: ME
- Focus: Long text task capability
- Numbers of Evaluation Categories/Subcategories: 1/-
- Evaluation Category: Long text task
CLongEval 2024-3 | All | ZH | HG & CI & MC | Paper | Github | Conjunto de datos
- Publisher: The Chinese University of Hong Kong et al.
- Size: 7267 instances
- License: MIT
- Question Type: SQ
- Evaluation Method: CE
- Focus: Long text task capability
- Numbers of Evaluation Categories/Subcategories: 7/-
- Evaluation Category: Long story QA, Long conversation memory, Long story summarization, Stacked news labeling, Stacked typo detection, Key-passage retrieval, Table querying
Counting-Stars 2024-3 | All | ZH | HG | Paper | Github | Dataset
- Publisher: Tencent MLPD
- Tamaño: -
- License: -
- Question Type: SQ
- Evaluation Method: CE
- Focus: Long text task capability
- Numbers of Evaluation Categories/Subcategories: 1/-
- Evaluation Category: Long text task
NeedleBench 2024-7 | All | EN & ZH | HG & CI | Paper | Github
- Publisher: Shanghai AI Laboratory et al.
- Tamaño: -
- License: -
- Question Type: Multi
- Evaluation Method: CE
- Focus: Assess bilingual long-context capabilities
- Numbers of Evaluation Categories/Subcategories: 3/-
- Evaluation Category: Single-retrieval, Multi-retrieval, Multi-reasoning

Herramienta

API-Bank 2023-4 | All | EN & PL | HG & MC | Paper | Github
- Publisher: Alibaba DAMO Academy et al.
- Size: 264 dialogues
- License: MIT
- Question Type: SQ
- Evaluation Method: HE & CE
- Focus: Plan step-by-step API calls, retrieve relevant APIs, and correctly execute API calls to meet human needs
- Numbers of Evaluation Categories/Subcategories: 3/-
- Evaluation Category: Call, Retrieval+Call, Plan+Retrieval+Call
APIBench 2023-5 | All | EN & PL | HG & MC | Paper | Github | Dataset | Sitio web
- Publisher: UC Berkeley et al.
- Size: 16450 instances
- License: Apache-2.0
- Question Type: SQ
- Evaluation Method: CE
- Focus: The reasoning ability for calling APIs
- Numbers of Evaluation Categories/Subcategories: 1/-
- Evaluation Category: API call
ToolBench 2023-5 | All | EN | HG & CI | Paper | Github
- Publisher: SambaNova Systems et al.
- Size: 795 instances
- License: Apache-2.0
- Question Type: SQ
- Evaluation Method: CE
- Focus: The enhancement in tool manipulation for real-world software tasks
- Numbers of Evaluation Categories/Subcategories: 8/-
- Evaluation Category: Open weather, The cat API, Home search, Trip booking, Google sheets, Virtual home, Web shop, Tabletop
ToolEyes 2024-1 | All | EN | HG | Paper | Github | Datasets
- Publisher: Fudan University
- Size: 382 instances
- License: Apache-2.0
- Question Type: SQ
- Evaluation Method: CE & ME
- Focus: The LLMs' tool learning capabilities in authentic scenarios
- Numbers of Evaluation Categories/Subcategories: 7/41
- Evaluation Category: Text generation, Data understanding, Real-time search, Application manipulation, Personal life, Information retrieval, Financial transactions

Agente

AgentBench 2023-8 | All | EN | HG & CI & MC | Paper | Github | Sitio web
- Publisher: Tsinghua University et al.
- Size: 1360 instances
- License: -
- Question Type: SQ
- Evaluation Method: CE
- Focus: LLM-as-Agent's reasoning and decision-making abilities
- Numbers of Evaluation Categories/Subcategories: 8/-
- Evaluation Category: Operating system, Database, Knowledge graph, Digital card game, Lateral thinking puzzles, House-holding, Web shopping, Web browsing
SuperCLUE-Agent 2023-10 | Not | ZH | - | Github | Website1 | Website2
- Publisher: CLUEbenchmark
- Tamaño: -
- License: -
- Question Type: SQ
- Evaluation Method: -
- Focus: Agent capabilities of LLMs
- Numbers of Evaluation Categories/Subcategories: 3/10
- Evaluation Category: Tool utilization, Task planning, Long-term and short-term memory

Código

BIRD 2023-5 | All | EN & PL | HG & CI & MC | Paper | Github | Dataset | Sitio web
- Publisher: The University of Hong Kong et al.
- Size: 12751 instances
- License: CC-BY-NC-4.0
- Question Type: SQ
- Evaluation Method: CE
- Focus: Text-to-SQL parsing
- Numbers of Evaluation Categories/Subcategories: 1/-
- Evaluation Category: Text-SQL
CodeXGLUE 2021-2 | All | EN & PL | CI | Paper | Github | Dataset
- Publisher: Peking University et al.
- Size: 4.12M instances
- License: C-UDA
- Question Type: SQ
- Evaluation Method: CE
- Focus: Program understanding and generation tasks
- Numbers of Evaluation Categories/Subcategories: 4/10
- Evaluation Category: Code-Code, Text-Code, Code-Text, Text-to-Text
DS-1000 2022-11 | All | EN & PL | HG | Paper | Github | Dataset | Sitio web
- Publisher: The University of Hong Kong et al.
- Size: 1000 instances
- License: CC-BY-SA-4.0
- Question Type: SQ
- Evaluation Method: CE
- Focus: Code generation
- Numbers of Evaluation Categories/Subcategories: 1/-
- Evaluation Category: Code generation
HumanEval 2021-7 | All | EN & PL | HG | Paper | Github
- Publisher: OpenAI et al.
- Size: 164 instances
- License: MIT
- Question Type: SQ
- Evaluation Method: CE
- Focus: The correctness of problem-solving abilities in the context of program synthesis
- Numbers of Evaluation Categories/Subcategories: 1/-
- Evaluation Category: Code generation
HumanEvalPack 2023-8 | All | EN & PL | HG & CI | Paper | Github | Dataset
- Publisher: Bigcode
- Size: 984 instances
- License: MIT
- Question Type: SQ
- Evaluation Method: CE
- Focus: The correctness of problem-solving abilities in the context of program synthesis
- Numbers of Evaluation Categories/Subcategories: 3/-
- Evaluation Category: HumanEvalFix, HumanEvalExplain, HumanEvalSynthesize
MTPB 2022-3 | All | EN & PL | HG | Paper | Github | Dataset
- Publisher: Salesforce Research
- Size: 115 instances
- License: Apache-2.0
- Question Type: SQ
- Evaluation Method: CE
- Focus: Multi-turn Programming
- Numbers of Evaluation Categories/Subcategories: 1/-
- Evaluation Category: Code generation
ODEX 2022-12 | All | Multi & PL | HG & CI | Paper | Github
- Publisher: Carnegie Mellon University et al.
- Size: 945 instances
- License: CC-BY-SA-4.0
- Question Type: SQ
- Evaluation Method: CE
- Focus: Natural language to Python code generation
- Numbers of Evaluation Categories/Subcategories: 1/-
- Evaluation Category: Code generation
APPS 2021-5 | All | EN & PL | HG | Paper | Github | Dataset
- Publisher: UC Berkeley et al.
- Size: 10000 instances
- License: MIT
- Question Type: SQ
- Evaluation Method: CE
- Focus: The ability to take an arbitrary natural language specification and generate satisfactory Python code
- Numbers of Evaluation Categories/Subcategories: 1/-
- Evaluation Category: Code generation
DomainEval 2024-8 | All | EN & PL | HG & CI & MC | Paper | Github | Dataset | Sitio web
- Publisher: Chinese Academy of Sciences et al.
- Size: 5892 instances
- License: -
- Question Type: SQ
- Evaluation Method: CE
- Focus: Evaluate LLMs' coding capabilities thoroughly
- Numbers of Evaluation Categories/Subcategories: 6/-
- Evaluation Category: Computation, Network, Basic operation, System, Visualization, Cryptography

OOD

BOSS 2023-6 | All | EN | CI | Paper | Github
- Publisher: Tsinghua University et al.
- Tamaño: -
- License: MIT
- Question Type: SQ
- Evaluation Method: CE
- Focus: Assess model performance under distribution shifts
- Numbers of Evaluation Categories/Subcategories: 5/20
- Evaluation Category: Sentiment analysis, Toxicity detection, Natural language inference, Named entity recognition, Extractive Question answering
GLUE-X 2022-11 | All | EN | CI | Paper | Github | Conjunto de datos
- Publisher: Westlake University et al.
- Size: 6404940 instances
- License: -
- Question Type: SQ
- Evaluation Method: CE
- Focus: The out-of-distribution (OOD) robustness
- Numbers of Evaluation Categories/Subcategories: 7/16
- Evaluation Category: Sentiment analysis, Linguistic acceptability, Textual similarity, Natural language inference, Question answering NLI, Textual entailment, Paraphrase

Ley

LAiW 2023-10 | Partial | ZH | CI | Paper | Github
- Publisher: Sichuan University et al.
- Tamaño: -
- License: -
- Question Type: SQ
- Evaluation Method: CE
- Focus: Legal capabilities
- Numbers of Evaluation Categories/Subcategories: 3/13
- Evaluation Category: Basic legal NLP, Basic legal application, Complex legal application
LawBench 2023-9 | All | ZH | HG & CI | Paper | Github | Dataset
- Publisher: Nanjing University et al.
- Tamaño: -
- License: Apache-2.0
- Question Type: Multi
- Evaluation Method: CE
- Focus: Legal capabilities
- Numbers of Evaluation Categories/Subcategories: 3/20
- Evaluation Category: Legal knowledge memorization, Legal knowledge understanding, Legal knowledge applying
LegalBench 2023-8 | All | EN | HG & CI | Paper | Github | Dataset | Sitio web
- Publisher: Stanford University et al.
- Size: 90417 instances
- License: -
- Question Type: SQ
- Evaluation Method: HE & CE
- Focus: Legal reasoning
- Numbers of Evaluation Categories/Subcategories: 6/162
- Evaluation Category: Issue-spotting, Rule-recall, Rule-application, Rule-conclusion, Interpretation, Rhetorical-understanding
LexGLUE 2021-10 | All | EN | CI | Paper | Github
- Publisher: University of Copenhagen et al.
- Size: 237014 instances
- License: -
- Question Type: SQ
- Evaluation Method: CE
- Focus: Legal capabilities
- Numbers of Evaluation Categories/Subcategories: 3/-
- Evaluation Category: Multi-label classification, Multi-class classification, Multiple choice QA
LEXTREME 2023-1 | All | Multi (24) | CI | Paper | Github
- Publisher: University of Bern et al.
- Size: 3508603 instances
- License: -
- Question Type: SQ
- Evaluation Method: CE
- Focus: Legal capabilities
- Numbers of Evaluation Categories/Subcategories: 18/-
- Evaluation Category: Brazilian court decisions, German argument mining, Greek legal code, Swiss judgment prediction, etc.
SCALE 2023-6 | All | Multi (5) | HG & CI | Papel | Dataset
- Publisher: University of Bern et al.
- Size: 1.86M instances
- License: CC-BY-SA
- Question Type: SQ
- Evaluation Method: CE
- Focus: Legal multidimensional abilities
- Numbers of Evaluation Categories/Subcategories: 4/-
- Evaluation Category: Processing long documents, Utilizing domain specific knowledge, Multilingual understanding, Multitasking
ArabLegalEval 2024-8 | All | AR | HG & CI & MC | Paper | Github | Dataset
- Publisher: THIQAH et al.
- Size: 37853 instances
- License: -
- Question Type: Multi
- Evaluation Method: ME
- Focus: Assess the Arabic legal knowledge of LLMs
- Numbers of Evaluation Categories/Subcategories: 3/-

Médico

CBLUE 2022-5 | All | ZH | HG & CI | Paper | Github
- Publisher: Zhejiang University et al.
- Size: 195820 instances
- License: Apache-2.0
- Question Type: SQ
- Evaluation Method: CE
- Focus: Chinese biomedical language understanding
- Numbers of Evaluation Categories/Subcategories: 5/8
- Evaluation Category: Information extraction from the medical text, normalization of the medical term, medical text classification, medical sentence similarity estimation, medical QA
CMB 2023-8 | All | ZH | HG | Paper | Github | Dataset | Sitio web
- Publisher: The Chinese University of Hong Kong et al.
- Size: 281047 instances
- License: Apache-2.0
- Question Type: Multi
- Evaluation Method: HE & CE & ME
- Focus: The performance of LLMs in the field of medicine
- Numbers of Evaluation Categories/Subcategories: 2/7
- Evaluation Category: CMB-Exam, CMB-Clin
HuaTuo26M-test 2023-5 | All | ZH | CI | Paper | Github | Dataset
- Publisher: The Chinese University of Hong Kong et al.
- Size: 6000 instances
- License: Apache-2.0
- Question Type: SQ
- Evaluation Method: CE
- Focus: Understand and generate complex medical language
- Numbers of Evaluation Categories/Subcategories: 3/-
- Evaluation Category: Medical consultant records, Encyclopedias, Knowledge bases
MultiMedQA 2022-12 | All | EN | HG & CI | Paper | Dataset
- Publisher: Google Research et al.
- Size: 212822 instances
- License: -
- Question Type: Multi
- Evaluation Method: HE & CE
- Focus: The performance in medical and clinical applications
- Numbers of Evaluation Categories/Subcategories: 1/-
- Evaluation Category: Medical question answering
PromptCBLUE 2023-4 | All | ZH | CI | Github
- Publisher: East China Normal University et al.
- Size: 20640 instances
- License: -
- Question Type: SQ
- Evaluation Method: CE
- Focus: The performance in Chinese medical scenarios
- Numbers of Evaluation Categories/Subcategories: 16/-
- Evaluation Category: Medical named entity recognition, Medical entity relation extraction, Medical event extraction, etc.
QiZhenGPT_eval 2023-5 | All | ZH | HG | Github | Dataset
- Publisher: Zhejiang University et al.
- Size: 94 instances
- License: GPL-3.0
- Question Type: SQ
- Evaluation Method: HE
- Focus: Indications for use of drugs
- Numbers of Evaluation Categories/Subcategories: 1/-
- Evaluation Category: Drug indication question answering
CLUE 2024-4 | Partical | EN | HG & CI & MC | Paper | Github
- Publisher: University Hospital Essen et al.
- Tamaño: -
- License: -
- Question Type: SQ
- Evaluation Method: CE
- Focus: Real-world clinical tasks
- Numbers of Evaluation Categories/Subcategories: 6/-
- Evaluation Category: MeDiSumQA, MeDiSumCode, MedNLI, MeQSum, Problem Summary, LongHealth

Financiero

BBF-CFLEB 2023-2 | All | ZH | HG & CI | Paper | Github | Sitio web
- Publisher: Fudan University et al.
- Size: 11327 instances
- License: -
- Question Type: SQ
- Evaluation Method: CE
- Focus: Language understanding and generation tasks in Chinese financial natural language processing
- Numbers of Evaluation Categories/Subcategories: 6/-
- Evaluation Category: FinNL, FinNA, FinRE, FinFE, FinQA, FinNSP
FinancelQ 2023-9 | All | ZH | HG & MC | Github
- Publisher: Du Xiaoman
- Size: 7173 instances
- License: CC-BY-NC-SA-4.0
- Question Type: OQ
- Evaluation Method: CE
- Focus: The knowledge and reasoning abilities in financial contexts
- Numbers of Evaluation Categories/Subcategories: 10/36
- Evaluation Category: Bank, Fund, Securities, Futures and derivatives, CICE, Actuarial science, Financial planning, CPA, Taxation, Economics
FinEval 2023-8 | All | ZH | HG | Paper | Github | Dataset | Sitio web
- Publisher: Shanghai University of Finance and Economics
- Size: 4661 instances
- License: CC-BY-NC-SA-4.0
- Question Type: OQ
- Evaluation Method: CE
- Focus: The performance in the financial domain knowledge
- Numbers of Evaluation Categories/Subcategories: 4/34
- Evaluation Category: Finance, Economy, Accounting, Certificate
FLUE 2022-10 | All | EN | CI | Paper | Sitio web
- Publisher: Georgia Institute of Technology et al.
- Size: 26292 instances
- License: -
- Question Type: SQ
- Evaluation Method: CE
- Focus: NLP tasks in the financial domain
- Numbers of Evaluation Categories/Subcategories: 5/6
- Evaluation Category: Financial sentiment analysis, News headline classification, Named entity recognition, Structure boundary detection, Question answering
FinBen 2024-2 | All | EN | CI | Paper | Github
- Publisher: The Fin AI et al.
- Size: 69805 instances
- License: -
- Question Type: SQ
- Evaluation Method: CE
- Focus: NLP tasks in the financial domain
- Numbers of Evaluation Categories/Subcategories: 3/6
- Evaluation Category: Foundamental tasks, Advanced cognitive engagement, General intelligence

Social Norms

CrowS-Pairs 2020-11 | All | EN | HG & CI | Paper | Github
- Publisher: New York University
- Size: 1508 instances
- License: CC-SA-4.0
- Question Type: SQ
- Evaluation Method: CE
- Focus: The presence of cultural biases and stereotypes in pretrained language models
- Numbers of Evaluation Categories/Subcategories: 9/-
- Evaluation Category: Race, Gender, Sexual orientation, Religion, Age, Nationality, Disability, Physical appearance, Occupation
SafetyBench 2023-9 | All | EN & ZH | HG & CI & MC | Paper | Github | Dataset | Sitio web
- Publisher: Tsinghua University et al.
- Size: 11435 instances
- License: MIT
- Question Type: OQ
- Evaluation Method: CE
- Focus: The safety of LLMs
- Numbers of Evaluation Categories/Subcategories: 7/-
- Evaluation Category: Offensiveness, Unfairness and bias, Physical health, Mental Health, Illegal activities, Ethics and morality, Privacy and Property
Safety-Prompts 2023-4 | Partial | ZH | MC | Paper | Github | Dataset | Sitio web
- Publisher: Tsinghua University et al.
- Size: 100K instances
- License: Apache-2.0
- Question Type: SQ
- Evaluation Method: HE & ME
- Focus: The safety of LLMs
- Numbers of Evaluation Categories/Subcategories: 2/13
- Evaluation Category: Typical security scenarios, Instruction attack
SuperCLUE-Safety 2023-9 | Not | ZH | - | Github | Sitio web
- Publisher: CLUEbenchmark
- Size: 4912 instances
- License: -
- Question Type: SQ
- Evaluation Method: ME
- Focus: The safety of LLMs
- Numbers of Evaluation Categories/Subcategories: 3/20+
- Evaluation Category: Traditional security category, Responsible artificial intelligence, Instruction attacks
TRUSTGPT 2023-6 | All | EN | CI | Paper | Github
- Publisher: Sichuan University et al.
- Size: 2000 instances
- License: MIT
- Question Type: SQ
- Evaluation Method: CE
- Focus: The performance in toxicity, bias, and value alignment
- Numbers of Evaluation Categories/Subcategories: 3/-
- Evaluation Category: Toxicity, Bias, Value-alignment

Factuality

FACTOR 2023-7 | Partial | EN | HG & CI & MC | Paper | Github
- Publisher: AI21 Labs
- Size: 4030 instances
- License: MIT
- Question Type: OQ
- Evaluation Method: CE
- Focus: The factuality of LLMs
- Numbers of Evaluation Categories/Subcategories: 2/-
- Evaluation Category: Wiki, News
FActScore 2023-5 | All | EN | HG & MC | Paper | Github
- Publisher: University of Washington et al.
- Size: 500 instances
- License: MIT
- Question Type: SQ
- Evaluation Method: HE & ME
- Focus: The factuality of LLMs
- Numbers of Evaluation Categories/Subcategories: 7/-
- Evaluation Category: Single-sentence contradiction (words or beyond words), Page-level contradiction, Subjective, Fact is irrelevant, Wiki is inconsistent & wrong, Annotation error
FactualityPrompt 2022-6 | All | EN | CI | Paper | Github
- Publisher: Hong Kong University of Science and Technology et al.
- Size: 16000 instances
- License: Apache-2.0
- Question Type: SQ
- Evaluation Method: CE
- Focus: The factuality of LLMs
- Numbers of Evaluation Categories/Subcategories: 2/-
- Evaluation Category: Factual prompts, Nonfactual prompts
FreshQA 2023-10 | All | EN | HG | Paper | Github
- Publisher: Google et al.
- Size: 600 instances
- License: -
- Question Type: SQ
- Evaluation Method: HE
- Focus: The factuality of LLMs
- Numbers of Evaluation Categories/Subcategories: 4/-
- Evaluation Category: Never-changing, Slow-changing, Fast-changing, False-premise
HalluQA 2023-10 | All | ZH | HG & MC | Paper | Github
- Publisher: Fudan University et al.
- Size: 450 instances
- License: -
- Question Type: SQ
- Evaluation Method: ME
- Focus: The factuality of LLMs
- Numbers of Evaluation Categories/Subcategories: 3/-
- Evaluation Category: Misleading, Misleading-hard, Knowledge
HaluEval 2023-5 | All | EN | HG & CI & MC | Paper | Github | Conjunto de datos
- Publisher: Renmin University of China et al.
- Size: 35000 instances
- License: MIT
- Question Type: SQ
- Evaluation Method: CE
- Focus: The factuality of LLMs
- Numbers of Evaluation Categories/Subcategories: 3/-
- Evaluation Category: QA, Dialogue, Summarization
TruthfulQA 2022-5 | All | EN | HG | Paper | Github
- Publisher: University of Oxford et al.
- Size: 817 instances
- License: Apache-2.0
- Question Type: SQ
- Evaluation Method: CE & ME
- Focus: The factuality of LLMs
- Numbers of Evaluation Categories/Subcategories: 38/-
- Evaluation Category: Health, Law, Conspiracies, Fiction, Misconceptions, Paranormal, Economics, Biology, Language, Indexical etc.
UHGEval 2023-11 | All | ZH | HG & MC | Paper | Github | Dataset
- Publisher: Renmin University of China et al.
- Size: 5141 instances
- License: Apache-2.0
- Question Type: Multi
- Evaluation Method: CE
- Focus: The factuality of LLMs
- Numbers of Evaluation Categories/Subcategories: 3/4
- Evaluation Category: Discriminative, Selective, Generative
HaluEval-Wild 2024-3 | Not | EN | HG & CI & MC | Papel
- Publisher: Carnegie Mellon University
- Size: 500 instances
- License: -
- Question Type: SQ
- Evaluation Method: -
- Focus: The factuality of LLMs
- Numbers of Evaluation Categories/Subcategories: 5/-
- Evaluation Category: Out-of-scope information, Complex reasoning, Inappropriate content, Beyond-modality interaction, Confused / Erroneous queries
RealTime QA 2022-7 | All | EN | HG | Paper | Github | Dataset | Sitio web
- Publisher: Toyota Technological Institute et al.
- Tamaño: -
- License: -
- Question Type: Multi
- Evaluation Method: CE
- Focus: The factuality of LLMs
- Numbers of Evaluation Categories/Subcategories: 1/-
- Evaluation Category: Latest knowledge Q&A
ChineseFactEval 2023-9 | All | ZH | HG & MC | Github | Dataset | Sitio web
- Publisher: Shanghai Jiao Tong University et al.
- Size: 125 instances
- License: Apache-2.0
- Question Type: SQ
- Evaluation Method: -
- Focus: The factuality of LLMs
- Numbers of Evaluation Categories/Subcategories: 7/-
- Evaluation Category: General domain, Scientific research, Medical, Law, Finance, Math, Chinese modern history
HalluDial 2024-6 | All | EN | CI & MC | Paper | Github | Dataset
- Publisher: BAAI et al.
- Size: 146856 instances
- License: -
- Question Type: SQ
- Evaluation Method: HE & CE & ME
- Focus: Automatic dialogue-level hallucination evaluation
- Numbers of Evaluation Categories/Subcategories: 2/-
- Evaluation Category: Hallucination detection, Hallucination localization and explanation

Evaluación

FairEval 2023-5 | All | EN | CI | Paper | Github | Dataset
- Publisher: Peking University et al.
- Size: 80 instances
- License: -
- Question Type: SQ
- Evaluation Method: CE
- Focus: The performance in determining the quality of output content from different models
- Numbers of Evaluation Categories/Subcategories: 1/-
- Evaluation Category: Evaluate the quality of answers
LLMEval2 2023-8 | All | Multi | CI | Papel | Github | Conjunto de datos
- Publisher: Chinese Academy of Sciences et al.
- Size: 2533 instances
- License: MIT
- Question Type: SQ
- Evaluation Method: CE
- Focus: The performance in determining the quality of output content from different models
- Numbers of Evaluation Categories/Subcategories: 1/-
- Evaluation Category: Evaluate the quality of answers
PandaLM_testset 2023-4 | All | EN | HG & MC | Papel | Github
- Publisher: Peking University et al.
- Size: 999 instances
- License: Apache-2.0
- Question Type: SQ
- Evaluation Method: CE
- Focus: The performance in determining the quality of output content from different models
- Numbers of Evaluation Categories/Subcategories: 1/-
- Evaluation Category: Evaluate the quality of answers

Multitarea

BBH 2022-10 | All | EN | CI | Papel | Github
- Publisher: Google Research et al.
- Size: 6511 instances
- License: MIT
- Question Type: Multi
- Evaluation Method: CE
- Focus: Challenging tasks that have proven difficult for prior language model evaluations
- Numbers of Evaluation Categories/Subcategories: 23/27
- Evaluation Category: Boolean expressions, Causal judgement, Date understanding, Disambiguation QA, etc.
BIG-Bench 2022-6 | All | Multi | HG & CI | Paper | Github
- Publisher: Google et al.
- Tamaño: -
- License: Apache-2.0
- Question Type: Multi
- Evaluation Method: CE
- Focus: The capabilities and limitations of language models
- Numbers of Evaluation Categories/Subcategories: 95/204
- Evaluation Category: Linguistics, Child development, Mathematics, Common sense reasoning, Biology, etc.
CLEVA 2023-8 | All | ZH | HG & CI | Paper | Github | Sitio web
- Publisher: The Chinese University of Hong Kong et al.
- Size: 370K instances
- License: CC-BY-NC-ND-4.0
- Question Type: SQ
- Evaluation Method: CE
- Focus: The performance of LLMs across various dimensions
- Numbers of Evaluation Categories/Subcategories: 2/31
- Evaluation Category: Ability, Application
CLiB 2023-6 | All | ZH | - | Github
- Publisher: jeinlee1991
- Size: 90 instances
- License: -
- Question Type: SQ
- Evaluation Method: HE
- Focus: Multidimensional capabilities
- Numbers of Evaluation Categories/Subcategories: 4/-
- Evaluation Category: Classification, Information extraction, Reading comprehension, Tabular question answering
decaNLP 2018-6 | All | EN | CI | Paper | Github
- Publisher: Salesforce Research
- Size: 2010693 instances
- License: BSD-3-Clause
- Question Type: SQ
- Evaluation Method: CE
- Focus: Multitask natural language processing capabilities
- Numbers of Evaluation Categories/Subcategories: 10/-
- Evaluation Category: Question answering, Machine translaion, Summarization, Natural language inference, Sentiment analysis, Semantic role labeling, Zero-shot relation extraction, Goal-oriented dialogue, Semantic parsing, Pronoun resolution
FlagEval 2023-6 | Partial | EN & ZH | HG & CI | Github | Sitio web
- Publisher: BAAI et al.
- Size: 84433 instances
- License: -
- Question Type: Multi
- Evaluation Method: HE & CE
- Focus: Multi-domain, multi-dimensional capabilities
- Numbers of Evaluation Categories/Subcategories: 3/21
- Evaluation Category: Choice qa, Classification, Generation qa
HELM 2022-11 | All | EN | CI | Paper | Github | Sitio web
- Publisher: Stanford University et al.
- Tamaño: -
- License: Apache-2.0
- Question Type: SQ
- Evaluation Method: HE & CE
- Focus: Evaluate LLMs on a wide range of scenarios and metrics
- Numbers of Evaluation Categories/Subcategories: 73/-
- Evaluation Category: Question answering, Information retrieval, Sentiment analysis, Toxicity detection, Aspirational scenarios, etc.
LLMEVAL-1 2023-5 | All | ZH | HG | Github
- Publisher: Fudan University et al.
- Size: 453 instances
- License: -
- Question Type: SQ
- Evaluation Method: HE & ME
- Focus: Multidimensional capabilities
- Numbers of Evaluation Categories/Subcategories: 17/-
- Evaluation Category: Fact-based question answering, Reading comprehension, Framework generation, Paragraph rewriting, etc.
LMentry 2023-7 | All | EN | HG | Paper | Github
- Publisher: Tel Aviv University et al.
- Size: 110703 instances
- License: -
- Question Type: SQ
- Evaluation Method: CE
- Focus: The performance on challenging tasks
- Numbers of Evaluation Categories/Subcategories: 25/-
- Evaluation Category: Sentence containing word, Sentence not containing word, Word containing letter, Word not containing letter, etc.
AlignBench 2023-11 | All | ZH | HG & MC | Papel | Github | Dataset
- Publisher: Tsinghua University et al.
- Size: 683 instances
- License: -
- Question Type: Multi
- Evaluation Method: ME
- Focus: Evaluate the alignment of LLMs on Chinese multitasks.
- Numbers of Evaluation Categories/Subcategories: 8/-
- Evaluation Category: Fundamental language ability, Advanced Chinese understanding, Open-ended questions, Writing ability, Logical reasoning, Mathematics, Task-oriented role play,
- Professional knowledge

Plurilingüe

XNLI 2018-10 | All | Multi (15) | HG | Paper | Github
- Publisher: Facebook AI et al.
- Size: 112500 instances
- License: CC-BY-NC-4.0
- Question Type: SQ
- Evaluation Method: CE
- Focus: Multilingual NLI
- Numbers of Evaluation Categories/Subcategories: 1/-
- Evaluation Category: Multilingual natural language inference
XTREME 2020-3 | All | Multi (40) | CI | Paper | Github | Sitio web
- Publisher: Carnegie Mellon University et al.
- Tamaño: -
- License: Apache-2.0
- Question Type: SQ
- Evaluation Method: CE
- Focus: The cross-lingual generalization capabilities
- Numbers of Evaluation Categories/Subcategories: 4/9
- Evaluation Category: Classification, Structured prediction, QA, Retrieval
MGSM 2022-10 | All | Multi (10) | CI | Paper | Github | Dataset
- Publisher: Google Research et al.
- Size: 2580 instances
- License: CC-BY-SA-4.0
- Question Type: SQ
- Evaluation Method: CE
- Focus: Multilingual mathematical reasoning abilities
- Numbers of Evaluation Categories/Subcategories: 1/-
- Evaluation Category: Math

Otro

EcomGPT_eval 2023-8 | All | EN & ZH | CI | Paper | Github
- Publisher: Alibaba
- Size: 6000 instances
- License: -
- Question Type: SQ
- Evaluation Method: CE
- Focus: E-commerce-related tasks
- Numbers of Evaluation Categories/Subcategories: 4/12
- Evaluation Category: Classification, Generation, Extraction, Others
- Domain: E-commerce
FewCLUE 2021-7 | Partial | ZH | CI | Paper | Github | Sitio web
- Publisher: CLUE team
- Size: 9 datasets
- License: -
- Question Type: SQ
- Evaluation Method: CE
- Focus: Compare different few-shot learning methods
- Numbers of Evaluation Categories/Subcategories: 3/9
- Evaluation Category: Single sentence tasks, Sentence pair tasks, Reading comprehension
- Domain: Few-shot learning
GeoBench 2023-6 | All | EN | HG | Paper | Github
- Publisher: Shanghai Jiao Tong University et al.
- Size: 2517 instances
- License: Apache-2.0
- Question Type: Multi
- Evaluation Method: HE & CE & ME
- Focus: LLMs' performance in understanding and utilizing geoscience knowledge
- Numbers of Evaluation Categories/Subcategories: 2/-
- Evaluation Category: NPEE, APTest
- Domain: Geoscience
Owl-Bench 2023-9 | All | EN & ZH | HG | Paper | Github
- Publisher: Beihang University et al.
- Size: 1317 instances
- License: -
- Question Type: Multi
- Evaluation Method: ME
- Focus: The performance in IT-related tasks
- Numbers of Evaluation Categories/Subcategories: 9/-
- Evaluation Category: Information security, Application, System architecture, Software architecture, Middleware, Network, Operating system, Infrastructure, Database
- Domain: IT
MINT 2023-9 | All | EN | CI | Paper | Github | Dataset | Sitio web
- Publisher: University of Illinois Urbana-Champaign et al.
- Size: 586 instances
- License: Apache-2.0
- Question Type: SQ
- Evaluation Method: CE
- Focus: Solve complex tasks through multi-turn interactions using tools and leveraging natural language feedback
- Numbers of Evaluation Categories/Subcategories: 3/-
- Evaluation Category: Code generation, Decision making, Reasoning
- Domain: Multi-turn interactions
PromptBench 2023-6 | All | EN | CI | Paper | Github
- Publisher: Microsoft Research et al.
- Size: 583884 instances
- License: MIT
- Question Type: SQ
- Evaluation Method: CE
- Focus: The models' robustness
- Numbers of Evaluation Categories/Subcategories: 10/15
- Evaluation Category: Sentiment analysis, Grammar correctness, Duplicate sentence detection, Natural language inference, etc.
- Domain: Robustness
EmotionBench 2023-8 | All | EN | HG & MC | Paper | Github
- Publisher: The Chinese University of Hong Kong et al.
- Tamaño: -
- License: GPL-3.0
- Question Type: SQ
- Evaluation Method: CE
- Focus: The empathy ability
- Numbers of Evaluation Categories/Subcategories: 8/36
- Evaluation Category: Anger, Anxiety, Depression, Frustration, Jealous, Guilt, Fear, Embarrassment
- Domain: Sentiment

Evaluation Platform

CLUE Benchmark Series
- SuperCLUE-Agent
- SuperCLUE-Auto
- SuperCLUE-Math6
- SuperCLUE-Safety
- SuperCLUE-Code3
- SuperCLUE-Video
- SuperCLUE-RAG
- SuperCLUE-Industry
- SuperCLUE-Role
OpenLLM Leaderboard
OpenCompass
MTEB Leaderboard
C-MTEB Leaderboard

Traditional NLP Datasets

Diverging from instruction fine-tuning datasets, we categorize text datasets dedicated to natural language tasks before the widespread adoption of LLMs as traditional NLP datasets.

Dataset information format：

 - Dataset name  Release Time | Language | Paper | Github | Dataset | Website
  - Publisher:
  - Train/Dev/Test/All Size: 
  - License:
  - Number of Entity Categories: (NER Task)
  - Number of Relationship Categories: (RE Task)

Question Answering

The task of question-answering requires the model to utilize its knowledge and reasoning capabilities to respond to queries based on provided text (which may be optional) and questions.

Comprensión de lectura

The task of reading comprehension entails presenting a model with a designated text passage and associated questions, prompting the model to understand the text for the purpose of answering the questions.

Selection & Judgment

BoolQ 2019-5 | EN | Papel | Github
- Publisher: University of Washington et al.
- Train/Dev/Test/All Size: 9427/3270/3245/15942
- License: CC-SA-3.0
CosmosQA 2019-9 | EN | Paper | Github | Dataset | Sitio web
- Publisher: University of Illinois Urbana-Champaign et al.
- Train/Dev/Test/All Size: 25588/3000/7000/35588
- License: CC-BY-4.0
CondaQA 2022-11 | EN | Paper | Github | Dataset
- Publisher: Carnegie Mellon University et al.
- Train/Dev/Test/All Size: 5832/1110/7240/14182
- License: Apache-2.0
PubMedQA 2019-9 | EN | Paper | Github | Dataset | Sitio web
- Publisher: University of Pittsburgh et al.
- Train/Dev/Test/All Size: -/-/-/273.5K
- License: MIT
MultiRC 2018-6 | EN | Paper | Github | Dataset
- Publisher: University of Pennsylvania et al.
- Train/Dev/Test/All Size: -/-/-/9872
- License: MultiRC License
RACE 2017-4 | EN | Paper | Dataset | Sitio web
- Publisher: Carnegie Mellon University
- Train/Dev/Test/All Size: 87866/4887/4934/97687
- License: -
C3 2019-4 | ZH | Paper | Github | Sitio web
- Publisher: Cornell University et al.
- Train/Dev/Test/All Size: 11869/3816/3892/19577
- License: -
ReClor 2020-2 | EN | Paper | Sitio web
- Publisher: National University of Singapore
- Train/Dev/Test/All Size: 4638/500/1000/6138
- License: -
DREAM 2020-2 | EN | Paper | Github | Sitio web
- Publisher: National University of Singapore
- Train/Dev/Test/All Size: 4638/500/1000/6138
- License: -
QuAIL 2020-4 | EN | Paper | Sitio web
- Publisher: University of Massachusetts Lowell
- Train/Dev/Test/All Size: 10346/-/2164/12510
- License: CC-NC-SA-4.0
DuReader Yes/No 2019-12 | ZH | Github1 | Github2
- Publisher: Baidu Inc. et al.
- Train/Dev/Test/All Size: 75K/5.5K/11K/91.5K
- License: Apache-2.0
MCTest 2013-10 | EN | Papel | Dataset
- Publisher: Microsoft Research
- Train/Dev/Test/All Size: 1200/200/600/2000
- License: -

Cloze Test

ChID 2019-6 | ZH | Paper | Github | Dataset
- Publisher: Tsinghua University et al.
- Train/Dev/Test/All Size: 605k/23.2K/83.3K/711.5K
- License: Apache-2.0
LAMBADA 2016-6 | EN | Paper | Dataset | Sitio web
- Publisher: University of Trento et al.
- Train/Dev/Test/All Size: 2662/4869/5153/12684
- License: CC-BY-4.0
CLOTH 2018-10 | EN | Paper | Dataset
- Publisher: Carnegie Melon University
- Train/Dev/Test/All Size: 76850/11067/11516/99433
- License: MIT
CMRC2019 2020-12 | ZH | Paper | Github | Sitio web
- Publisher: Harbin Institute of Technology et al.
- Train/Dev/Test/All Size: 100009/3053/5118/108180
- License: CC-BY-SA-4.0

Answer Extraction

SQuAD 2016-11 | EN | Paper | Dataset
- Publisher: Stanford University
- Train/Dev/Test/All Size: 87599/10570/9533/107702
- License: CC-BY-4.0
SQuAD 2.0 2018-6 | EN | Paper | Dataset
- Publisher: Stanford University
- Train/Dev/Test/All Size: 130319/11873/8862/151054
- License: CC-BY-SA-4.0
HOTPOTQA 2018-9 | EN | Paper | Dataset | Sitio web
- Publisher: Carnegie Mellon University et al.
- Train/Dev/Test/All Size: 90447/7405/7405/105257
- License: CC-BY-SA-4.0
TriviaQA 2017-7 | EN | Paper | Github | Conjunto de datos
- Publisher: Univ. of Washington et al.
- Train/Dev/Test/All Size: -/-/-/95000
- License: Apache-2.0
Natural Questions 2019-X | EN | Papel | Github | Dataset
- Publisher: Google Research
- Train/Dev/Test/All Size: 307372/7830/7842/323044
- License: CC-BY-4.0
ReCoRD 2018-10 | EN | Paper | Sitio web
- Publisher: Johns Hopkins University et al.
- Train/Dev/Test/All Size: 100730/10000/10000/120730
- License: -
QuAC 2018-8 | EN | Paper | Dataset | Sitio web
- Publisher: AI2 et al.
- Train/Dev/Test/All Size: 83568/7354/7353/98407
- License: CC-BY-SA-4.0
TyDiQA 2020-3 | Multi (11) | Paper | Github | Dataset
- Publisher: Google Research
- Train/Dev/Test/All Size: 116916/18670/18751/154337
- License: Apache-2.0
CMRC2018 2019-11 | ZH | Paper | Github
- Publisher: Harbin Institute of Technology et al.
- Train/Dev/Test/All Size: 10321/3351/4895/18567
- License: CC-BY-SA-4.0
Adversarial QA 2020-2 | EN | Paper | Github | Dataset
- Publisher: University College London
- Train/Dev/Test/All Size: 30000/3000/3000/36000
- License: MIT
Quoref 2019-8 | EN | Paper | Sitio web
- Publisher: AI2 et al.
- Train/Dev/Test/All Size: 19399/2418/2537/24354
- License: CC-BY-4.0
MLQA 2020-7 | Multi (7) | Paper | Github | Dataset
- Publisher: Facebook AI Research et al.
- Train/Dev/Test/All Size: -/4199/42246/46445
- License: CC-BY-SA-3.0
DuReader Robust 2020-3 | ZH | Paper | Github1 | Github2
- Publisher: Baidu Inc. et al.
- Train/Dev/Test/All Size: 15K/1.4K/4.8K/21.2K
- License: Apache-2.0
DuReader Checklist 2021-3 | ZH | Github1 | Github2
- Publisher: Baidu Inc. et al.
- Train/Dev/Test/All Size: 3K/1.1K/4.5K/8.6K
- License: Apache-2.0
CUAD 2021-3 | EN | Papel | Dataset
- Publisher: UC Berkeley et al.
- Train/Dev/Test/All Size: 22450/-/4182/26632
- License: CC-BY-4.0
MS MARCO 2016-11 | EN | Paper | Github | Dataset
- Publisher: Microsoft AI & Research
- Train/Dev/Test/All Size: 808731/101093/101092/1010916
- License: MIT

Unrestricted QA

DROP 2019-6 | EN | Paper | Sitio web
- Publisher: University of California et al.
- Train/Dev/Test/All Size: 77409/9536/9622/96567
- License: CC-BY-4.0
CoQA 2018-8 | EN | Paper | Sitio web
- Publisher: Stanford University
- Train/Dev/Test/All Size: -/-/-/127K
- License: -
QASPER 2021-5 | EN | Paper | Sitio web
- Publisher: AI2 et al.
- Train/Dev/Test/All Size: -/-/-/5049
- License: CC-BY-4.0
DuoRC 2018-7 | EN | Paper | Dataset | Sitio web
- Publisher: IBM Research et al.
- Train/Dev/Test/All Size: 130261/27914/27914/186089
- License: MIT
DuReader 2.0 2018-4 | ZH | Paper | Github1 | Github2
- Publisher: Baidu Inc. et al.
- Train/Dev/Test/All Size: -/-/-/200K
- License: Apache-2.0

Knowledge QA

In the knowledge QA task, models respond to questions by leveraging world knowledge, common sense, scientific insights, domain-specific information, and more.

ARC 2018-3 | EN | Paper | Sitio web
- Publisher: AI2
- Train/Dev/Test/All Size: 3370/869/3548/7787
- License: CC-BY-SA
CommonsenseQA 2018-11 | EN | Paper | Github | Dataset | Sitio web
- Publisher: Tel-Aviv University et al.
- Train/Dev/Test/All Size: 9797/1225/1225/12247
- License: MIT
OpenBookQA 2018-10 | EN | Paper | Github | Conjunto de datos
- Publisher: AI2 et al.
- Train/Dev/Test/All Size: 4957/500/500/5957
- License: Apache-2.0
PIQA 2019-11 | EN | Paper | Github | Dataset
- Publisher: AI2 et al.
- Train/Dev/Test/All Size: 16.1K/1.84K/3.08K/21.02K
- License: MIT
JEC-QA 2019-11 | EN | Paper | Github | Dataset | Sitio web
- Publisher: Tsinghua University et al.
- Train/Dev/Test/All Size: -/-/26365/26365
- License: CC-NC-ND-4.0
CMD 2019-X | ZH | Github | Dataset
- Publisher: Toyhom
- Train/Dev/Test/All Size: -/-/-/792099
- License: MIT
cMedQA2 2018-11 | ZH | Paper | Dataset
- Publisher: National University of Defense Technology
- Train/Dev/Test/All Size: 100000/4000/4000/108000
- License: GPL-3.0
HEAD-QA 2019-7 | EN & ES | Paper | Github | Dataset | Sitio web
- Publisher: Universidade da Coruna
- Train/Dev/Test/All Size: 2657/1366/2742/13530
- License: MIT
SciQ 2017-9 | EN | Paper | Dataset | Sitio web
- Publisher: University College London et al.
- Train/Dev/Test/All Size: 11679/1000/1000/13679
- License: CC-BY-NC-3.0
WikiQA 2015-9 | EN | Paper | Dataset | Sitio web
- Publisher: Georgia Institute of Technology et al.
- Train/Dev/Test/All Size: 2118/296/633/3047
- License: Microsoft Research Data License
ECQA 2021-8 | EN | Paper | Github
- Publisher: IIT Delhi et al.
- Train/Dev/Test/All Size: 7598/1090/2194/10882
- License: CDLA-Sharing-1.0
PsyQA 2021-6 | ZH | Paper | Github
- Publisher: The CoAI group et al.
- Train/Dev/Test/All Size: -/-/-/22346
- License: PsyQA User Agreement
WebMedQA 2018-12 | ZH | Papel | Github
- Publisher: Chinese Academy of Sciences et al.
- Train/Dev/Test/All Size: 50610/6337/6337/63284
- License: Apache-2.0
WebQuestions 2013-10 | EN | Paper | Dataset
- Publisher: Stanford University
- Train/Dev/Test/All Size: 3778/-/2032/5810
- License: -

Reasoning QA

The focal point of reasoning QA tasks is the requirement for models to apply abilities such as logical reasoning, multi-step inference, and causal reasoning in answering questions.

STRATEGYQA 2021-1 | EN | Paper | Sitio web
- Publisher: Tel Aviv University et al.
- Train/Dev/Test/All Size: 2290/-/490/2780
- License: MIT
COPA 2011-6 | EN | Paper | Sitio web
- Publisher: Indiana University et al.
- Train/Dev/Test/All Size: -/500/500/1000
- License: BSD 2-Clause
HellaSwag 2019-7 | EN | Paper | Github
- Publisher: University of Washington et al.
- Train/Dev/Test/All Size: 39905/10042/10003/59950
- License: MIT
StoryCloze 2016-6 | EN | Paper | Conjunto de datos
- Publisher: University of Rochester et al.
- Train/Dev/Test/All Size: -/1871/1871/3742
- License: -
Social IQa 2019-4 | EN | Paper | Dataset
- Publisher: AI2
- Train/Dev/Test/All Size: 33410/1954/-/35364
- License: -
LogiQA 2020-7 | EN & ZH | Paper | Github
- Publisher: Fudan University et al.
- Train/Dev/Test/All Size: 7376/651/651/8678
- License: -
PROST 2021-8 | EN | Paper | Github | Dataset
- Publisher: University of Colorado Boulder
- Train/Dev/Test/All Size: -/-/18736/18736
- License: Apache-2.0
QuaRTz 2019-11 | EN | Paper | Dataset | Sitio web
- Publisher: AI2
- Train/Dev/Test/All Size: 2696/384/784/3864
- License: CC-BY-4.0
WIQA 2019-9 | EN | Paper | Dataset | Sitio web
- Publisher: AI2
- Train/Dev/Test/All Size: 29808/6894/3993/40695
- License: -
QASC 2019-10 | EN | Paper | Dataset | Sitio web
- Publisher: AI2 et al.
- Train/Dev/Test/All Size: 8134/926/920/9980
- License: CC-BY-4.0
QuaRel 2018-11 | EN | Paper | Sitio web
- Publisher: AI2
- Train/Dev/Test/All Size: 1941/278/552/2771
- License: CC-BY-4.0
ROPES 2019-8 | EN | Paper | Dataset | Sitio web
- Publisher: AI2
- Train/Dev/Test/All Size: 10K/1.6K/1.7K/13.3K
- License: CC-BY-4.0
CREAK 2021-9 | EN | Paper | Github
- Publisher: The University of Texas at Austin
- Train/Dev/Test/All Size: 10176/1371/1371/13418
- License: MIT

Recognizing Textual Entailment

The primary objective of tasks related to Recognizing Textual Entailment (RTE) is to assess whether information in one textual segment can be logically inferred from another.

ANLI 2019-10 | EN | Paper | Github | Dataset
- Publisher: UNC Chapel Hill et al.
- Train/Dev/Test/All Size: 162865/3200/3200/169265
- License: CC-NC-4.0
RTE - | EN | Paper1 | Paper2 | Paper3 | Paper4 | Dataset
- Publisher: The PASCAL Recognising Textual Entailment Challenge
- Train/Dev/Test/All Size: 2.49K/277/3K/5.77K
- License: CC-BY-4.0
WANLI 2022-1 | EN | Papel | Dataset
- Publisher: University of Washington et al.
- Train/Dev/Test/All Size: 102885/-/5000/107885
- License: CC-BY-4.0
MedNLI 2018-8 | EN | Paper | Github | Dataset | Sitio web
- Publisher: University of Massachusetts Lowell et al.
- Train/Dev/Test/All Size: 11232/1395/1422/14049
- License: -
CommitmentBank 2019-X | EN | Paper | Github | Dataset
- Publisher: The Ohio State University et al.
- Train/Dev/Test/All Size: -/-/-/1200
- License: -
MultiNLI 2018-6 | EN | Paper | Dataset
- Publisher: New York University
- Train/Dev/Test/All Size: 392702/19647/-/412349
- License: -
SNLI 2015-8 | EN | Paper | Dataset
- Publisher: Stanford Linguistics et al.
- Train/Dev/Test/All Size: 550152/10000/10000/570152
- License: CC-BY-SA-4.0
OCNLI 2020-10 | ZH | Paper | Github
- Publisher: Indiana University et al.
- Train/Dev/Test/All Size: 50K/3K/3K/56K
- License: CC-BY-NC-2.0
CMNLI 2020-12 | ZH | Github | Dataset
- Publisher: CLUE team
- Train/Dev/Test/All Size: 391783/12426/13880/418089
- License: -
CINLID 2021-4 | ZH | Dataset
- Publisher: Gao et al.
- Train/Dev/Test/All Size: 80124/-/26708/106832
- License: -

Matemáticas

Mathematical assignments commonly involve standard mathematical calculations, theorem validations, and mathematical reasoning tasks, among others.

GSM8K 2021-10 | EN | Paper | Github | Dataset
- Publisher: OpenAI
- Train/Dev/Test/All Size: 7.5K/-/1K/8.5K
- License: MIT
SVAMP 2021-3 | EN | Paper | Github
- Publisher: Microsoft Research India
- Train/Dev/Test/All Size: -/-/-/1000
- License: MIT
ASDiv 2021-6 | EN | Paper | Github | Dataset
- Publisher: Institute of Information Science
- Train/Dev/Test/All Size: -/-/-/2305
- License: CC-BY-NC-4.0
MATH 2021-3 | EN | Paper | Github | Dataset
- Publisher: UC Berkeley et al.
- Train/Dev/Test/All Size: 7500/-/5000/12500
- License: MIT
Ape210K 2020-9 | ZH | Paper | Github
- Publisher: Yuanfudao AI Lab et al.
- Train/Dev/Test/All Size: 200488/5000/5000/210488
- License: -
Math23K 2017-9 | ZH | Paper | Github
- Publisher: Tencent AI Lab
- Train/Dev/Test/All Size: -/-/-/23161
- License: MIT
MathQA 2019-5 | EN | Paper | Dataset | Sitio web
- Publisher: University of Washington et al.
- Train/Dev/Test/All Size: 29837/4475/2985/37297
- License: Apache-2.0
AQUA-RAT 2017-7 | EN | Paper | Github | Dataset
- Publisher: DeepMind
- Train/Dev/Test/All Size: 100949/250/250/101499
- License: Apache-2.0
NaturalProofs 2021-4 | EN | Paper | Github
- Publisher: University of Washington et al.
- Train/Dev/Test/All Size: -/-/-/80795
- License: MIT

Coreference Resolution

The core objective of tasks related to coreference resolution is the identification of referential relationships within texts.

WSC 2012-X | EN | Paper | Dataset
- Publisher: University of Toronto et al.
- Train/Dev/Test/All Size: -/-/285/285
- License: CC-BY-4.0
DPR 2012-7 | EN | Paper | Dataset
- Publisher: University of Texas at Dallas
- Train/Dev/Test/All Size: 1322/-/564/1886
- License: -
WinoGrande 2019-7 | EN | Paper | Github | Dataset
- Publisher: AI2 et al.
- Train/Dev/Test/All Size: 63238/1267/1767/66272
- License: CC-BY
WiC 2018-8 | EN | Paper | Sitio web
- Publisher: University of Cambridge
- Train/Dev/Test/All Size: 5428/638/1400/7466
- License: CC-NC-4.0
WinoWhy 2020-7 | EN | Paper | Github
- Publisher: HKUST
- Train/Dev/Test/All Size: -/-/-/43972
- License: MIT
CLUEWSC2020 2020-12 | ZH | Paper | Github1 | Github2
- Publisher: CLUE team
- Train/Dev/Test/All Size: 1244/304/290/1838
- License: -

Análisis de sentimientos

The sentiment analysis task, commonly known as emotion classification, seeks to analyze and deduce the emotional inclination of provided texts, commonly categorized as positive, negative, or neutral sentiments.

IMDB 2011-6 | EN | Paper | Dataset
- Publisher: Stanford University
- Train/Dev/Test/All Size: 25000/-/25000/50000
- License: -
Sentiment140 2009-X | EN | Paper | Dataset
- Publisher: Stanford University
- Train/Dev/Test/All Size: 1600000/-/359/1600359
- License: -
SST-2 2013-10 | EN | Paper | Dataset
- Publisher: Stanford University
- Train/Dev/Test/All Size: 67349/872/1821/70042
- License: -
EPRSTMT 2021-7 | ZH | Paper | Github
- Publisher: CLUE team
- Train/Dev/Test/All Size: 32/32/1363/20992
- License: -

Semantic Matching

The task of semantic matching entails evaluating the semantic similarity or degree of correspondence between two sequences of text.

MRPC 2005-X | EN | Papel
- Publisher: Microsoft Research
- Train/Dev/Test/All Size: 4076/-/1725/5801
- License: -
QQP 2018-11 | EN | Papel | Dataset
- Publisher: New York University et al.
- Train/Dev/Test/All Size: 364K/-/-/364K
- License: -
PAWS 2019-6 | EN | Paper | Github | Dataset
- Publisher: Google AI Language
- Train/Dev/Test/All Size: 49401/8000/8000/65401
- License: -
STSB 2017-8 | Multi (10) | Paper | Github | Dataset | Sitio web
- Publisher: Google Research et al.
- Train/Dev/Test/All Size: 5749/1500/1379/8628
- License: -
AFQMC 2020-12 | ZH | Papel
- Publisher: CLUE team
- Train/Dev/Test/All Size: 34.3K/4.3K/3.9K/42.5K
- License: -
BQ 2018-10 | ZH | Paper | Dataset
- Publisher: Harbin Institute of Technology et al.
- Train/Dev/Test/All Size: 100000/10000/10000/120000
- License: -
LCQMC 2018-8 | ZH | Papel
- Publisher: Harbin Institute of Technology et al.
- Train/Dev/Test/All Size: 238766/8802/12500/260068
- License: CC-BY-4.0
PAWS-X 2019-8 | Multi (6) | Paper | Github
- Publisher: Google Research
- Train/Dev/Test/All Size: 296406/11815/11844/320065
- License: -
BUSTM 2021-7 | ZH | Paper | Github
- Publisher: CLUE team
- Train/Dev/Test/All Size: 32/32/3772/8087
- License: -
DuQM 2021-9 | ZH | Paper | Github1 | Github2
- Publisher: Baidu Inc. et al.
- Train/Dev/Test/All Size: -/-/-/10121
- License: Apache-2.0

Text Generation

The narrow definition of text generation tasks is bound by provided content and specific requirements. It involves utilizing benchmark data, such as descriptive terms and triplets, to generate corresponding textual descriptions.

CommonGen 2019-11 | EN | Paper | Github | Dataset
- Publisher: University of Southern California et al.
- Train/Dev/Test/All Size: 67389/4018/1497/72904
- License: MIT
DART 2020-7 | EN | Paper | Github | Dataset
- Publisher: Yale University et al.
- Train/Dev/Test/All Size: 30526/2768/6959/40253
- License: MIT
E2E 2017-6 | EN | Paper | Github | Dataset
- Publisher: Heriot-Watt University
- Train/Dev/Test/All Size: 42061/4672/4693/51426
- License: CC-BY-SA-3.0
WebNLG 2017-7 | EN & RU | Paper | Github | Dataset
- Publisher: LORIA et al.
- Train/Dev/Test/All Size: 49665/6490/7930/64085
- License: CC-BY-NC-SA-4.0

Traducción de texto

Text translation involves transforming text from one language to another.

WMT - | Multi | Sitio web
- Publisher: ACL et al.
- Train/Dev/Test/All Size: -/-/-/-
- License: -
NLLB 2022-7 | Multi | Paper | Github
- Publisher: NLLB Team et al.
- Train/Dev/Test/All Size: -/-/-/-
- License: MIT
IWSLT 2017 2017-12 | Multi (11) | Paper | Dataset | Sitio web
- Publisher: FBK et al.
- Train/Dev/Test/All Size: 1108475/4442/41921/1154838
- License: CC-BY-NC-ND-4.0

Text Summarization

The task of text summarization pertains to the extraction or generation of a brief summary or headline from an extended text to encapsulate its primary content.

AESLC 2019-7 | EN | Paper | Github | Dataset
- Publisher: Yale University et al.
- Train/Dev/Test/All Size: 14436/1960/1906/18302
- License: CC-BY-NC-SA-4.0
CNN-DM 2017-4 | EN | Paper | Dataset
- Publisher: Stanford University et al.
- Train/Dev/Test/All Size: 287113/13368/11490/311971
- License: Apache-2.0
MultiNews 2019-7 | EN | Paper | Github | Dataset
- Publisher: Yale University
- Train/Dev/Test/All Size: 44972/5622/5622/56216
- License: -
Newsroom 2018-6 | EN | Paper | Dataset
- Publisher: Cornell University
- Train/Dev/Test/All Size: 995041/108837/108862/1212740
- License: -
SAMSum 2019-11 | EN | Paper | Dataset
- Publisher: Cornell University
- Train/Dev/Test/All Size: 14732/818/819/16369
- License: CC-BY-NC-ND-4.0
XSum 2018-10 | EN | Paper | Github | Dataset
- Publisher: University of Edinburgh
- Train/Dev/Test/All Size: 204045/11332/11334/226711
- License: MIT
Opinion Abstracts 2016-6 | EN | Papel | Dataset
- Publisher: Northeastern University et al.
- Train/Dev/Test/All Size: 5990/-/-/5990
- License: -
WikiLingua 2020-10 | Multi (18) | Paper | Github | Dataset
- Publisher: Columbia University et al.
- Train/Dev/Test/All Size: -/-/-/770087
- License: CC-BY-3.0
LCSTS 2015-6 | ZH | Paper | Dataset
- Publisher: Harbin Institute of Technology
- Train/Dev/Test/All Size: 2400000/10000/1000/2411000
- License: CC-BY-4.0
CNewSum 2021-10 | ZH | Paper | Github | Dataset | Sitio web
- Publisher: ByteDance
- Train/Dev/Test/All Size: 275596/14356/14355/304307
- License: Apache-2.0
XL-Sum 2021-8 | Multi (45) | Paper | Dataset
- Publisher: BUET et al.
- Train/Dev/Test/All Size: 1122857/114198/114198/1351253
- License: CC-BY-NC-SA-4.0
WikiHow 2018-10 | EN | Paper | Github
- Publisher: University of California
- Train/Dev/Test/All Size: -/-/-/230K
- License: CC-BY-NC-SA
MediaSum 2021-3 | EN | Paper | Github | Dataset
- Publisher: Microsoft Cognitive Services Research Group
- Train/Dev/Test/All Size: 443596/10000/10000/463596
- License: -

Text Classification

Text classification tasks aim to assign various text instances to predefined categories, comprising text data and category labels as pivotal components.

AGNEWS 2015-9 | EN | Paper | Dataset | Sitio web
- Publisher: New York University
- Train/Dev/Test/All Size: 120000/-/7600/127600
- License: -
TNEWS 2020-11 | ZH | Paper | Github | Dataset
- Publisher: CLUE team
- Train/Dev/Test/All Size: 53.3K/10K/10K/73.3K
- License: -
IFLYTEK 2020-12 | ZH | Papel
- Publisher: CLUE team
- Train/Dev/Test/All Size: 12.1K/2.6K/2.6K/17.3K
- License: -
MARC 2020-11 | Multi (6) | Paper | Dataset
- Publisher: Amazon et al.
- Train/Dev/Test/All Size: 1200000/30000/30000/1260000
- License: -
THUCNews 2016-X | ZH | Github | Sitio web
- Publisher: Tsinghua University
- Train/Dev/Test/All Size: -/-/-/1672165
- License: MIT
CSLDCP 2021-7 | ZH | Paper | Github | Sitio web
- Publisher: CLUE team
- Train/Dev/Test/All Size: 536/536/4783/23966
- License: -

Text Quality Evaluation

The task of text quality evaluation, also referred to as text correction, involves the identification and correction of grammatical, spelling, or language usage errors in text.

CoLA 2018-5 | EN | Paper | Sitio web
- Publisher: New York University
- Train/Dev/Test/All Size: 8511/1043/-/9554
- License: CC-BY-4.0
SIGHAN - | ZH | Paper1 | Paper2 | Paper3 | Dataset1 | Dataset2 | Dataset3
- Publisher: Chaoyang Univ. of Technology et al.
- Train/Dev/Test/All Size: 6476/-/3162/9638
- License: -
YACLC 2021-12 | ZH | Papel | Github
- Publisher: Beijing Language and Culture University et al.
- Train/Dev/Test/All Size: 8000/1000/1000/10000
- License: -
CSCD-IME 2022-11 | ZH | Paper | Github
- Publisher: Tencent Inc
- Train/Dev/Test/All Size: 30000/5000/5000/40000
- License: MIT

Text-to-Code

The Text-to-Code task involves models converting user-provided natural language descriptions into computer-executable code, thereby achieving the desired functionality or operation.

MBPP 2021-8 | EN & PL | Paper | Github
- Publisher: Google Research
- Train/Dev/Test/All Size: -/-/974/974
- License: -
DuSQL 2020-11 | ZH & PL | Paper | Dataset
- Publisher: Baidu Inc. et al.
- Train/Dev/Test/All Size: 18602/2039/3156/23797
- License: -
CSpider 2019-11 | ZH & PL | Paper | Github | Sitio web
- Publisher: Westlake University
- Train/Dev/Test/All Size: -/-/-/10181
- License: CC-BY-SA-4.0
Spider 2018-9 | EN & PL | Paper | Github | Sitio web
- Publisher: Yale University
- Train/Dev/Test/All Size: -/-/-/10181
- License: CC-BY-SA-4.0

Named Entity Recognition

The Named Entity Recognition (NER) task aims to discern and categorize named entities within a given text.

WUNT2017 2017-9 | EN | Paper | Dataset
- Publisher: Johns Hopkins University et al.
- Train/Dev/Test/All Size: 3394/1009/1287/5690
- License: CC-BY-4.0
- Number of Entity Categories: 6
Few-NERD 2021-5 | EN | Paper | Github | Dataset | Sitio web
- Publisher: Tsinghua University et al.
- Train/Dev/Test/All Size: -/-/-/188200
- License: CC-BY-SA-4.0
- Number of Entity Categories: 66
CoNLL2003 2003-6 | EN & DE | Paper | Dataset
- Publisher: University of Antwerp
- Train/Dev/Test/All Size: 14041/3250/3453/20744
- License: -
- Number of Entity Categories: 4
OntoNotes 5.0 2013-10 | Multi (3) | Papel | Dataset | Sitio web
- Publisher: Boston Childrens Hospital and Harvard Medical School et al.
- Train/Dev/Test/All Size: 59924/8528/8262/76714
- License: -
- Number of Entity Categories: 18
MSRA 2006-7 | ZH | Paper | Dataset
- Publisher: University of Chicago
- Train/Dev/Test/All Size: 46364/-/4365/50729
- License: CC-BY-4.0
- Number of Entity Categories: 3
Youku NER 2019-6 | ZH | Paper | Github | Dataset
- Publisher: Singapore University of Technology and Design et al.
- Train/Dev/Test/All Size: 8001/1000/1001/10002
- License: -
- Number of Entity Categories: 9
Taobao NER 2019-6 | ZH | Paper | Github | Conjunto de datos
- Publisher: Singapore University of Technology and Design et al.
- Train/Dev/Test/All Size: 6000/998/1000/7998
- License: -
- Number of Entity Categories: 9
Weibo NER 2015-9 | ZH | Paper | Github | Dataset
- Publisher: Johns Hopkins University
- Train/Dev/Test/All Size: 1350/269/270/1889
- License: CC-BY-SA-3.0
- Number of Entity Categories: 4
CLUENER 2020-1 | ZH | Paper | Github | Dataset
- Publisher: CLUE Organization
- Train/Dev/Test/All Size: 10748/1343/1345/13436
- License: -
- Number of Entity Categories: 10
Resume 2018-7 | ZH | Paper | Github
- Publisher: Singapore University of Technology and Design
- Train/Dev/Test/All Size: 3821/463/477/4761
- License: -
- Number of Entity Categories: 8

Relation Extraction

The endeavor of Relation Extraction (RE) necessitates the identification of connections between entities within textual content. This process typically includes recognizing and labeling pertinent entities, followed by the determination of the specific types of relationships that exist among them.

Dialogue RE 2020-7 | EN & ZH | Paper | Github | Sitio web
- Publisher: Tencent AI Lab et al.
- Train/Dev/Test/All Size: 6100/2034/2034/10168
- License: -
- Number of Relationship Categories: 36
TACRED 2017-9 | EN | Paper | Dataset | Sitio web
- Publisher: Stanford University
- Train/Dev/Test/All Size: 68124/22631/15509/106264
- License: LDC
- Number of Relationship Categories: 42
DocRED 2019-7 | EN | Paper | Github
- Publisher: Tsinghua University et al.
- Train/Dev/Test/All Size: 1546589/12332/12842/1571763
- License: MIT
- Number of Relationship Categories: 96
FewRel 2018-10 | EN | Paper1 | Paper2 | Github | Sitio web
- Publisher: Tsinghua University
- Train/Dev/Test/All Size: -/-/-/70000
- License: CC-BY-SA-4.0
- Number of Relationship Categories: 100

Multitarea

Multitask datasets hold significance as they can be concurrently utilized for different categories of NLP tasks.

CSL 2022-9 | ZH | Papel | Github
- Publisher: School of Information Engineering et al.
- Train/Dev/Test/All Size: -/-/-/396209
- License: Apache-2.0
QED 2021-3 | EN | Paper | Github
- Publisher: Stanford University et al.
- Train/Dev/Test/All Size: 7638/1355/-/8993
- License: CC-BY-SA-3.0 & GFDL
METS-CoV 2022-9 | EN | Paper | Github
- Publisher: Zhejiang University et al.
- Train/Dev/Test/All Size: -/-/-/-
- License: Apache-2.0

Multi-modal Large Language Models (MLLMs) Datasets

Pre-training Corpora

Documentos

mOSCAR : A large-scale multilingual and multimodal document-level corpus
- Paper: mOSCAR: A Large-scale Multilingual and Multimodal Document-level Corpus
- Github: https://oscar-project.github.io/documentation/versions/mOSCAR/
OBELISC : An open web-scale filtered dataset of interleaved image-text documents
- Paper: OBELICS: An Open Web-Scale Filtered Dataset of Interleaved Image-Text Documents
- Github: https://github.com/huggingface/OBELICS
- Dataset: https://huggingface.co/datasets/HuggingFaceM4/OBELICS

Instruction Fine-tuning Datasets

Remote Sensing

MMRS-1M : Multi-sensor remote sensing instruction dataset
- Paper: EarthGPT: A Universal Multi-modal Large Language Model for Multi-sensor Image Comprehension in Remote Sensing Domain
- Github: https://github.com/wivizhang/EarthGPT

Images + Videos

VideoChat2-IT : Instruction fine-tuning dataset for images/videos
- Paper: MVBench: A Comprehensive Multi-modal Video Understanding Benchmark
- Dataset: https://huggingface.co/datasets/OpenGVLab/VideoChat2-IT

Visual Document Understanding

InstructDoc : A dataset for zero-shot generalization of visual document understanding
- Paper: InstructDoc: A Dataset for Zero-Shot Generalization of Visual Document Understanding with Instructions
- Github: https://github.com/nttmdlab-nlp/InstructDoc
- Dataset: https://github.com/nttmdlab-nlp/InstructDoc

General

ALLaVA-4V Data : The multimodal instruction fine-tuning dataset for the ALLaVA model
- Paper: ALLaVA: Harnessing GPT4V-synthesized Data for A Lite Vision-Language Model
- Github: https://github.com/FreedomIntelligence/ALLaVA
- Dataset: https://huggingface.co/datasets/FreedomIntelligence/ALLaVA-4V

Evaluation Datasets

Video Understanding

MVBench : A comprehensive multi-modal video understanding benchmark
- Paper: MVBench: A Comprehensive Multi-modal Video Understanding Benchmark
- Github: https://github.com/OpenGVLab/Ask-Anything/tree/main/video_chat2
- Dataset: https://huggingface.co/datasets/OpenGVLab/MVBench

Sujeto

OlympiadBench : An Olympiad-level bilingual multimodal scientific benchmark
- Paper: OlympiadBench: A Challenging Benchmark for Promoting AGI with Olympiad-Level Bilingual Multimodal Scientific Problems
- Github: https://github.com/OpenBMB/OlympiadBench
- Dataset: https://github.com/OpenBMB/OlympiadBench
MMMU : A new benchmark designed to evaluate MLLMs on massive multi-discipline tasks
- Paper: MMMU: A Massive Multi-discipline Multimodal Understanding and Reasoning Benchmark for Expert AGI
- Github: https://github.com/MMMU-Benchmark/MMMU
- Dataset: https://huggingface.co/datasets/MMMU/MMMU

Multitarea

MMT-Bench : A comprehensive multimodal benchmark for evaluating large vision-language models towards multitask AGI
- Paper: MMT-Bench: A Comprehensive Multimodal Benchmark for Evaluating Large Vision-Language Models Towards Multitask AGI
- Github: https://github.com/OpenGVLab/MMT-Bench
- Dataset: https://huggingface.co/datasets/Kaining/MMT-Bench

Long Input

MM-NIAH : The first benchmark specifically designed to systematically evaluate the capability of existing MLLMs to comprehend long multimodal documents
- Paper: Needle In A Multimodal Haystack
- Github: https://github.com/OpenGVLab/MM-NIAH
- Dataset: https://github.com/OpenGVLab/MM-NIAH

Factuality

MultiTrust : The first comprehensive and unified benchmark on the trustworthiness of MLLMs across five primary aspects: truthfulness, safety, robustness, fairness, and privacy
- Paper: Benchmarking Trustworthiness of Multimodal Large Language Models: A Comprehensive Study
- Github: https://github.com/thu-ml/MMTrustEval
- Website: https://multi-trust.github.io/#leaderboard

Médico

MultiMed : A benchmark designed to evaluate and enable large-scale learning across a wide spectrum of medical modalities and tasks
- Paper: MultiMed: Massively Multimodal and Multitask Medical Understanding
MedTrinity-25M : A large-scale multimodal dataset with multigranular annotations for medicine
- Paper: MedTrinity-25M: A Large-scale Multimodal Dataset with Multigranular Annotations for Medicine
- Github: https://github.com/UCSC-VLAA/MedTrinity-25M
- Dataset: https://huggingface.co/datasets/UCSC-VLAA/MedTrinity-25M
- Website: https://yunfeixie233.github.io/MedTrinity-25M/

Image Understanding

MMIU : A comprehensive evaluation suite designed to assess LVLMs across a wide range of multi-image tasks
- Paper: MMIU: Multimodal Multi-image Understanding for Evaluating Large Vision-Language Models
- Github: https://github.com/OpenGVLab/MMIU
- Dataset: https://huggingface.co/datasets/FanqingM/MMIU-Benchmark
- Website: https://mmiu-bench.github.io/

Retrieval Augmented Generation (RAG) Datasets

CRUD-RAG : A comprehensive Chinese benchmark for RAG
- Paper: CRUD-RAG: A Comprehensive Chinese Benchmark for Retrieval-Augmented Generation of Large Language Models
- Github: https://github.com/IAAR-Shanghai/CRUD_RAG
- Dataset: https://github.com/IAAR-Shanghai/CRUD_RAG
WikiEval : To do correlation analysis of difference metrics proposed in RAGAS
- Paper: RAGAS: Automated Evaluation of Retrieval Augmented Generation
- Github: https://github.com/explodinggradients/ragas
- Dataset: https://huggingface.co/datasets/explodinggradients/WikiEval
RGB : A benchmark for RAG
- Paper: Benchmarking Large Language Models in Retrieval-Augmented Generation
- Github: https://github.com/chen700564/RGB
- Dataset: https://github.com/chen700564/RGB
RAG-Instruct-Benchmark-Tester : An updated benchmarking test dataset for RAG use cases in the enterprise
- Dataset: https://huggingface.co/datasets/llmware/rag_instruct_benchmark_tester
- Website: https://medium.com/@darrenoberst/how-accurate-is-rag-8f0706281fd9
ARES : An automated evaluation framework for RAG
- Paper: ARES: An Automated Evaluation Framework for Retrieval-Augmented Generation Systems
- Github: https://github.com/stanford-futuredata/ARES
- Dataset: https://github.com/stanford-futuredata/ARES
ALCE : The quality assessment benchmark for context and responses
- Paper: Enabling Large Language Models to Generate Text with Citations
- Github: https://github.com/princeton-nlp/ALCE
- Dataset: https://huggingface.co/datasets/princeton-nlp/ALCE-data
CRAG : A comprehensive RAG benchmark
- Paper: CRAG -- Comprehensive RAG Benchmark
- Website: https://www.aicrowd.com/challenges/meta-comprehensive-rag-benchmark-kdd-cup-2024
RAGEval ：A framework for automatically generating evaluation datasets to evaluate the knowledge usage ability of different LLMs in different scenarios
- Paper: RAGEval: Scenario Specific RAG Evaluation Dataset Generation Framework
- Github: https://github.com/OpenBMB/RAGEval
- Dataset: https://github.com/OpenBMB/RAGEval
LFRQA ：A dataset of human-written long-form answers for cross-domain evaluation in RAG-QA systems
- Paper: RAG-QA Arena: Evaluating Domain Robustness for Long-form Retrieval Augmented Question Answering
- Github: https://github.com/awslabs/rag-qa-arena
MultiHop-RAG : Benchmarking retrieval-augmented generation for multi-hop queries
- Paper: MultiHop-RAG: Benchmarking Retrieval-Augmented Generation for Multi-Hop Queries
- Github: https://github.com/yixuantt/MultiHop-RAG/
- Dataset: https://huggingface.co/datasets/yixuantt/MultiHopRAG

Contacto

Información del contacto:

Lianwen Jin：[email protected]

Yang Liu：[email protected]

Due to our current limited human resources to manage such a vast amount of data resources, we regret that we are unable to include all data resources at this moment. If you find any important data resources that have not yet been included, we warmly invite you to submit relevant papers, data links, and other information to us. We will evaluate them, and if appropriate, we will include the data in the Awesome-LLMs-Datasets and the survey paper . Your assistance and support are greatly appreciated!

Citación

If you wish to cite this project, please use the following citation format:

 @article{liu2024survey,
  title={Datasets for Large Language Models: A Comprehensive Survey},
  author={Liu, Yang and Cao, Jiahuan and Liu, Chongyu and Ding, Kai and Jin, Lianwen},
  journal={arXiv preprint arXiv:2402.18041},
  year={2024}
}

Expandir