LLaMa2lang
v0.5
Este repositório contém scripts convenientes para ajustar o LLaMa3-8B (ou qualquer outro modelo básico) para bate-papo em qualquer idioma (que não seja o inglês). A justificativa por trás disso é que o LLaMa3 é treinado principalmente com dados em inglês e, embora funcione até certo ponto para outros idiomas, seu desempenho é fraco em comparação com o inglês.
Combine o poder do ajuste fino com o poder do RAG - confira nosso repositório RAG Me Up no RAG, que pode ser usado em cima de seus modelos ajustados com LLaMa2Lang.
pip install -r requirements.txt
# Translate OASST1 to target language
python translate.py m2m target_lang checkpoint_location
# Combine the checkpoint files into a dataset
python combine_checkpoints.py input_folder output_location
# Finetune
python finetune.py tuned_model dataset_name instruction_prompt
# Optionally finetune with DPO (RLHF)
python finetune_dpo.py tuned_model dataset_name instruction_prompt
# Run inference
python run_inference.py model_name instruction_prompt input
O processo que seguimos para ajustar um modelo básico como LLaMa3 para uma linguagem específica é o seguinte:
Os itens a seguir foram testados, mas potencialmente mais funcionarão
O processo acima pode ser totalmente executado em uma GPU Google Colab T4 gratuita. A última etapa, entretanto, só pode ser executada com êxito com janelas de contexto curtas o suficiente e um lote de no máximo 2. Além disso, a tradução na etapa 2 leva cerca de 36 horas no total para qualquer idioma, portanto deve ser executada em várias etapas se você quero ficar com uma GPU Google Colab gratuita.
Nossos modelos ajustados para a etapa 5 foram executados usando um A40 no vasto.ai e custaram menos de um dólar para cada modelo, sendo concluídos em cerca de 1,5 horas.
Certifique-se de que o pytorch esteja instalado e funcionando para o seu ambiente (preferencialmente o uso de CUDA): https://pytorch.org/get-started/locally/
Clone o repositório e instale os requisitos.
pip install -r requirements.txt
usage: translate.py [-h] [--quant8] [--quant4] [--base_dataset BASE_DATASET] [--base_dataset_text_field BASE_DATASET_TEXT_FIELD] [--base_dataset_lang_field BASE_DATASET_LANG_FIELD]
[--checkpoint_n CHECKPOINT_N] [--batch_size BATCH_SIZE] [--max_length MAX_LENGTH] [--cpu] [--source_lang SOURCE_LANG]
{opus,mbart,madlad,m2m,nllb,seamless_m4t_v2,towerinstruct} ... target_lang checkpoint_location
Translate an instruct/RLHF dataset to a given target language using a variety of translation models
positional arguments:
{opus,mbart,madlad,m2m,nllb,seamless_m4t_v2,towerinstruct}
The model/architecture used for translation.
opus Translate the dataset using HelsinkiNLP OPUS models.
mbart Translate the dataset using mBART.
madlad Translate the dataset using Google's MADLAD models.
m2m Translate the dataset using Facebook's M2M models.
nllb Translate the dataset using Facebook's NLLB models.
seamless_m4t_v2 Translate the dataset using Facebook's SeamlessM4T-v2 multimodal models.
towerinstruct Translate the dataset using Unbabel's Tower Instruct. Make sure your target language is in the 10 languages supported by the model.
target_lang The target language. Make sure you use language codes defined by the translation model you are using.
checkpoint_location The folder the script will write (JSONized) checkpoint files to. Folder will be created if it doesn't exist.
options:
-h, --help show this help message and exit
--quant8 Optional flag to load the translation model in 8 bits. Decreases memory usage, increases running time
--quant4 Optional flag to load the translation model in 4 bits. Decreases memory usage, increases running time
--base_dataset BASE_DATASET
The base dataset to translate, defaults to OpenAssistant/oasst1
--base_dataset_text_field BASE_DATASET_TEXT_FIELD
The base dataset's column name containing the actual text to translate. Defaults to text
--base_dataset_lang_field BASE_DATASET_LANG_FIELD
The base dataset's column name containing the language the source text was written in. Defaults to lang
--checkpoint_n CHECKPOINT_N
An integer representing how often a checkpoint file will be written out. To start off, 400 is a reasonable number.
--batch_size BATCH_SIZE
The batch size for a single translation model. Adjust based on your GPU capacity. Default is 10.
--max_length MAX_LENGTH
How much tokens to generate at most. More tokens might be more accurate for lengthy input but creates a risk of running out of memory. Default is unlimited.
--cpu Forces usage of CPU. By default GPU is taken if available.
--source_lang SOURCE_LANG
Source language to select from OASST based on lang property of dataset
Se você quiser mais parâmetros para os diferentes modelos de tradução, execute:
python translate.py [MODEL] -h
Certifique-se de especificar os parâmetros específicos do modelo antes de especificar os parâmetros comuns da lista acima. Exemplos de chamadas:
# Using M2M with 4bit quantization and differen batch sizes to translate Dutch
python translate.py m2m nl ./output_nl --quant4 --batch_size 20
# Using madlad 7B with 8bit quantization for German with different max_length
python translate.py madlad --model_size 7b de ./output_de --quant8 --batch_size 5 --max_length 512
# Be sure to use target language codes that the model you use understands
python translate.py mbart xh_ZA ./output_xhosa
python translate.py nllb nld_Latn ./output_nl
HF_TOKEN configurada conforme a documentação. usage: combine_checkpoints.py [-h] input_folder output_location
Combine checkpoint files from translation.
positional arguments:
input_folder The checkpoint folder used in translation, with the target language appended.
Example: "./output_nl".
output_location Where to write the Huggingface Dataset. Can be a disk location or a Huggingface
Dataset repository.
options:
-h, --help show this help message and exit
usage: finetune.py [-h] [--base_model BASE_MODEL] [--base_dataset_text_field BASE_DATASET_TEXT_FIELD] [--base_dataset_rank_field BASE_DATASET_RANK_FIELD] [--base_dataset_id_field BASE_DATASET_ID_FIELD] [--base_dataset_parent_field BASE_DATASET_PARENT_FIELD]
[--base_dataset_role_field BASE_DATASET_ROLE_FIELD] [--quant8] [--noquant] [--max_seq_length MAX_SEQ_LENGTH] [--num_train_epochs NUM_TRAIN_EPOCHS] [--batch_size BATCH_SIZE] [--threads_output_name THREADS_OUTPUT_NAME] [--thread_template THREAD_TEMPLATE]
[--padding PADDING]
tuned_model dataset_name instruction_prompt
Finetune a base instruct/chat model using (Q)LoRA and PEFT
positional arguments:
tuned_model The name of the resulting tuned model.
dataset_name The name of the dataset to use for fine-tuning. This should be the output of the combine_checkpoints script.
instruction_prompt An instruction message added to every prompt given to the chatbot to force it to answer in the target language. Example: "You are a generic chatbot that always answers in English."
options:
-h, --help show this help message and exit
--base_model BASE_MODEL
The base foundation model. Default is "NousResearch/Meta-Llama-3-8B-Instruct".
--base_dataset_text_field BASE_DATASET_TEXT_FIELD
The dataset's column name containing the actual text to translate. Defaults to text
--base_dataset_rank_field BASE_DATASET_RANK_FIELD
The dataset's column name containing the rank of an answer given to a prompt. Defaults to rank
--base_dataset_id_field BASE_DATASET_ID_FIELD
The dataset's column name containing the id of a text. Defaults to message_id
--base_dataset_parent_field BASE_DATASET_PARENT_FIELD
The dataset's column name containing the parent id of a text. Defaults to parent_id
--base_dataset_role_field BASE_DATASET_ROLE_FIELD
The dataset's column name containing the role of the author of the text (eg. prompter, assistant). Defaults to role
--quant8 Finetunes the model in 8 bits. Requires more memory than the default 4 bit.
--noquant Do not quantize the finetuning. Requires more memory than the default 4 bit and optional 8 bit.
--max_seq_length MAX_SEQ_LENGTH
The maximum sequence length to use in finetuning. Should most likely line up with your base model's default max_seq_length. Default is 512.
--num_train_epochs NUM_TRAIN_EPOCHS
Number of epochs to use. 2 is default and has been shown to work well.
--batch_size BATCH_SIZE
The batch size to use in finetuning. Adjust to fit in your GPU vRAM. Default is 4
--threads_output_name THREADS_OUTPUT_NAME
If specified, the threads created in this script for finetuning will also be saved to disk or HuggingFace Hub.
--thread_template THREAD_TEMPLATE
A file containing the thread template to use. Default is threads/template_fefault.txt
--padding PADDING What padding to use, can be either left or right.
6.1 [OPCIONAL] Ajuste fino usando DPO (semelhante ao RLHF)
usage: finetune_dpo.py [-h] [--base_model BASE_MODEL] [--base_dataset_text_field BASE_DATASET_TEXT_FIELD] [--base_dataset_rank_field BASE_DATASET_RANK_FIELD] [--base_dataset_id_field BASE_DATASET_ID_FIELD] [--base_dataset_parent_field BASE_DATASET_PARENT_FIELD] [--quant8]
[--noquant] [--max_seq_length MAX_SEQ_LENGTH] [--max_prompt_length MAX_PROMPT_LENGTH] [--num_train_epochs NUM_TRAIN_EPOCHS] [--batch_size BATCH_SIZE] [--threads_output_name THREADS_OUTPUT_NAME] [--thread_template THREAD_TEMPLATE] [--max_steps MAX_STEPS]
[--padding PADDING]
tuned_model dataset_name instruction_prompt
Finetune a base instruct/chat model using (Q)LoRA and PEFT using DPO (RLHF)
positional arguments:
tuned_model The name of the resulting tuned model.
dataset_name The name of the dataset to use for fine-tuning. This should be the output of the combine_checkpoints script.
instruction_prompt An instruction message added to every prompt given to the chatbot to force it to answer in the target language. Example: "You are a generic chatbot that always answers in English."
options:
-h, --help show this help message and exit
--base_model BASE_MODEL
The base foundation model. Default is "NousResearch/Meta-Llama-3-8B-Instruct".
--base_dataset_text_field BASE_DATASET_TEXT_FIELD
The dataset's column name containing the actual text to translate. Defaults to text
--base_dataset_rank_field BASE_DATASET_RANK_FIELD
The dataset's column name containing the rank of an answer given to a prompt. Defaults to rank
--base_dataset_id_field BASE_DATASET_ID_FIELD
The dataset's column name containing the id of a text. Defaults to message_id
--base_dataset_parent_field BASE_DATASET_PARENT_FIELD
The dataset's column name containing the parent id of a text. Defaults to parent_id
--quant8 Finetunes the model in 8 bits. Requires more memory than the default 4 bit.
--noquant Do not quantize the finetuning. Requires more memory than the default 4 bit and optional 8 bit.
--max_seq_length MAX_SEQ_LENGTH
The maximum sequence length to use in finetuning. Should most likely line up with your base model's default max_seq_length. Default is 512.
--max_prompt_length MAX_PROMPT_LENGTH
The maximum length of the prompts to use. Default is 512.
--num_train_epochs NUM_TRAIN_EPOCHS
Number of epochs to use. 2 is default and has been shown to work well.
--batch_size BATCH_SIZE
The batch size to use in finetuning. Adjust to fit in your GPU vRAM. Default is 4
--threads_output_name THREADS_OUTPUT_NAME
If specified, the threads created in this script for finetuning will also be saved to disk or HuggingFace Hub.
--thread_template THREAD_TEMPLATE
A file containing the thread template to use. Default is threads/template_fefault.txt
--max_steps MAX_STEPS
The maximum number of steps to run DPO for. Default is -1 which will run the data through fully for the number of epochs but this will be very time-consuming.
--padding PADDING What padding to use, can be either left or right.
6.1 [OPCIONAL] Ajuste fino usando ORPO (semelhante ao RLHF)
usage: finetune_orpo.py [-h] [--base_model BASE_MODEL] [--base_dataset_text_field BASE_DATASET_TEXT_FIELD] [--base_dataset_rank_field BASE_DATASET_RANK_FIELD] [--base_dataset_id_field BASE_DATASET_ID_FIELD] [--base_dataset_parent_field BASE_DATASET_PARENT_FIELD] [--quant8]
[--noquant] [--max_seq_length MAX_SEQ_LENGTH] [--max_prompt_length MAX_PROMPT_LENGTH] [--num_train_epochs NUM_TRAIN_EPOCHS] [--batch_size BATCH_SIZE] [--threads_output_name THREADS_OUTPUT_NAME] [--thread_template THREAD_TEMPLATE] [--max_steps MAX_STEPS]
[--padding PADDING]
tuned_model dataset_name instruction_prompt
Finetune a base instruct/chat model using (Q)LoRA and PEFT using ORPO (RLHF)
positional arguments:
tuned_model The name of the resulting tuned model.
dataset_name The name of the dataset to use for fine-tuning. This should be the output of the combine_checkpoints script.
instruction_prompt An instruction message added to every prompt given to the chatbot to force it to answer in the target language. Example: "You are a generic chatbot that always answers in English."
options:
-h, --help show this help message and exit
--base_model BASE_MODEL
The base foundation model. Default is "NousResearch/Meta-Llama-3-8B-Instruct".
--base_dataset_text_field BASE_DATASET_TEXT_FIELD
The dataset's column name containing the actual text to translate. Defaults to text
--base_dataset_rank_field BASE_DATASET_RANK_FIELD
The dataset's column name containing the rank of an answer given to a prompt. Defaults to rank
--base_dataset_id_field BASE_DATASET_ID_FIELD
The dataset's column name containing the id of a text. Defaults to message_id
--base_dataset_parent_field BASE_DATASET_PARENT_FIELD
The dataset's column name containing the parent id of a text. Defaults to parent_id
--quant8 Finetunes the model in 8 bits. Requires more memory than the default 4 bit.
--noquant Do not quantize the finetuning. Requires more memory than the default 4 bit and optional 8 bit.
--max_seq_length MAX_SEQ_LENGTH
The maximum sequence length to use in finetuning. Should most likely line up with your base model's default max_seq_length. Default is 512.
--max_prompt_length MAX_PROMPT_LENGTH
The maximum length of the prompts to use. Default is 512.
--num_train_epochs NUM_TRAIN_EPOCHS
Number of epochs to use. 2 is default and has been shown to work well.
--batch_size BATCH_SIZE
The batch size to use in finetuning. Adjust to fit in your GPU vRAM. Default is 4
--threads_output_name THREADS_OUTPUT_NAME
If specified, the threads created in this script for finetuning will also be saved to disk or HuggingFace Hub.
--thread_template THREAD_TEMPLATE
A file containing the thread template to use. Default is threads/template_fefault.txt
--max_steps MAX_STEPS
The maximum number of steps to run ORPO for. Default is -1 which will run the data through fully for the number of epochs but this will be very time-consuming.
--padding PADDING What padding to use, can be either left or right.
usage: run_inference.py [-h] model_name instruction_prompt input
Script to run inference on a tuned model.
positional arguments:
model_name The name of the tuned model that you pushed to Huggingface in the previous
step.
instruction_prompt An instruction message added to every prompt given to the chatbot to force
it to answer in the target language.
input The actual chat input prompt. The script is only meant for testing purposes
and exits after answering.
options:
-h, --help show this help message and exit
Como posso saber qual modelo de tradução escolher para meu idioma de destino?
Nós ajudamos você com nosso script benchmark.py que ajuda a fazer uma boa estimativa (o conjunto de dados que usamos é o mesmo no qual os modelos OPUS são treinados, então os resultados são sempre favoráveis ao OPUS). Para uso, consulte a ajuda deste script abaixo. Os modelos são carregados em quantização de 4 bits e executados em uma pequena amostra do subconjunto de livros OPUS.
Certifique-se de usar os idiomas que ocorrem mais comumente em seu conjunto de dados base como idioma_fonte e seu idioma de tradução de destino como idioma_alvo. Para OASST1, por exemplo, certifique-se de executar pelo menos en e es como idiomas de origem.
usage: benchmark.py [-h] [--cpu] [--start START] [--n N] [--max_length MAX_LENGTH] source_language target_language included_models
Benchmark all the different translation models for a specific source and target language to find out which performs best. This uses 4bit quantization to limit GPU usage. Note:
the outcomes are indicative - you cannot assume corretness of the BLEU and CHRF scores but you can compare models against each other relatively.
positional arguments:
source_language The source language you want to test for. Check your dataset to see which occur most prevalent or use English as a good start.
target_language The source language you want to test for. This should be the language you want to apply the translate script on. Note: in benchmark, we use 2-character
language codes, in constrast to translate.py where you need to specify whatever your model expects.
included_models Comma-separated list of models to include. Allowed values are: opus, m2m_418m, m2m_1.2b, madlad_3b, madlad_7b, madlad_10b, madlad_7bbt, mbart,
nllb_distilled600m, nllb_1.3b, nllb_distilled1.3b, nllb_3.3b, seamless
options:
-h, --help show this help message and exit
--cpu Forces usage of CPU. By default GPU is taken if available.
--start START The starting offset to include sentences from the OPUS books dataset from. Defaults to 0.
--n N The number of sentences to benchmark on. Defaults to 100.
--max_length MAX_LENGTH
How much tokens to generate at most. More tokens might be more accurate for lengthy input but creates a risk of running out of memory. Default is 512.
Já criamos e continuaremos a criar vários conjuntos de dados e modelos. Quer ajudar a democratizar os LLMs? Clone o repositório e crie conjuntos de dados e modelos para outras linguagens e, em seguida, crie um PR.
| Holandês UnderstandingLing/oasst1_nl | Espanhol UnderstandingLing/oast1_es | Francês CompreenderLing/oasst1_fr | Alemão UnderstandingLing/oasst1_de |
| Xaviviro catalão/oasst1_ca | Português UnderstandingLing/oast1_pt | Árabe HeshamHaroon/oast-árabe | Italiano UnderstandingLing/oasst1_it |
| Russo UnderstandingLing/oasst1_ru | Hindi EntendaLing/oasst1_hi | Chinês CompreenderLing/oasst1_zh | Cristãos poloneses/oasst1_pl |
| Japonês UnderstandingLing/oasst1_jap | Basco xezpeleta/oasst1_eu | Entenda BengaliLing/oasst1_bn | Turco UnderstandingLing/oasst1_tr |
Certifique-se de ter acesso ao modelo LLaMa3-8B do Meta e defina seu HF_TOKEN antes de usar esses modelos.
| UnderstandingLing/Llama-3-8B-Instruct-nl Holandês | UnderstandingLing/Llama-3-8B-Instruct-es Espanhol | UnderstandingLing/Llama-3-8B-Instruct-fr Francês | UnderstandingLing/Llama-3-8B-Instruct-de Alemão |
| UnderstandingLing/Llama-3-8B-Instruct-pt Português | UnderstandingLing/Llama-3-8B-Instruct-it Italiano | UnderstandingLing/Llama-3-8B-Instruct-hi Hindi | UnderstandingLing/Llama-3-8B-Instruct-ru Russo |
| Holandês UnderstandingLing/oasst1_nl_threads | Espanhol UnderstandingLing/oasst1_es_threads | Francês UnderstandingLing/oasst1_fr_threads | Alemão UnderstandingLing/oasst1_de_threads |
| Xaviviro catalão/oasst1_ca_threads | Português UnderstandingLing/oast1_pt_threads | Árabe HeshamHaroon/oasst-arabic_threads | Italiano UnderstandingLing/oasst1_it_threads |
| Russo UnderstandingLing/oasst1_ru_threads | Hindi UnderstandingLing/oasst1_hi_threads | Chinês UnderstandingLing/oasst1_zh_threads | Cristãos poloneses/oasst1_pl_threads |
| Japonês UnderstandingLing/oasst1_jap_threads | Basco xezpeleta/oasst1_eu_threads | Bengali UnderstandingLing/oasst1_bn_threads | Turco UnderstandingLing/oasst1_tr_threads |
| UnderstandingLing/llama-2-7b-chat-nl Holandês | UnderstandingLing/llama-2-7b-chat-es Espanhol | UnderstandingLing/llama-2-7b-chat-fr Francês | UnderstandingLing/llama-2-7b-chat-de Alemão |
| xaviviro/llama-2-7b-chat-ca Catalão | UnderstandingLing/llama-2-7b-chat-pt Português | HeshamHaroon/llama-2-7b-chat-ar Árabe | UnderstandingLing/llama-2-7b-chat-it Italiano |
| UnderstandingLing/llama-2-7b-chat-ru Russo | EntendaLing/llama-2-7b-chat-hi Hindi | EntendaLing/llama-2-7b-chat-zh Chinês | cristãos/llama-2-7b-chat-pl-polish-polski polonês |
| xezpeleta/llama-2-7b-chat-eu Basco | EntendaLing/llama-2-7b-chat-bn bengali | UnderstandingLing/llama-2-7b-chat-tr Turco |
| UnderstandingLing/Mistral-7B-Instruct-v0.2-nl Holandês | UnderstandingLing/Mistral-7B-Instruct-v0.2-es Espanhol | UnderstandingLing/Mistral-7B-Instruct-v0.2-de Alemão |
| UnderstandingLing/llama-2-13b-chat-nl Holandês | UnderstandingLing/llama-2-13b-chat-es Espanhol | UnderstandingLing/llama-2-13b-chat-fr Francês |
| UnderstandingLing/Mixtral-8x7B-Instruct-nl Holandês |
<s>[INST] <<SYS>> Je bent een generieke chatbot die altijd in het Nederlands antwoord geeft. <</SYS>> Wat is de hoofdstad van Nederland? [/INST] Amsterdam</s>
<s>[INST] <<SYS>> Je bent een generieke chatbot die altijd in het Nederlands antwoord geeft. <</SYS>> Wat is de hoofdstad van Nederland? [/INST] Amsterdam</s><s>[INST] Hoeveel inwoners heeft die stad? [/INST] 850 duizend inwoners (2023)</s>
<s>[INST] <<SYS>> Je bent een generieke chatbot die altijd in het Nederlands antwoord geeft. <</SYS>> Wat is de hoofdstad van Nederland? [/INST] Amsterdam</s><s>[INST] Hoeveel inwoners heeft die stad? [/INST] 850 duizend inwoners (2023)</s><s>[INST] In welke provincie ligt die stad? [/INST] In de provincie Noord-Holland</s>
<s>[INST] <<SYS>> Je bent een generieke chatbot die altijd in het Nederlands antwoord geeft. <</SYS>> Wie is de minister-president van Nederland? [/INST] Mark Rutte is sinds 2010 minister-president van Nederland. Hij is meerdere keren herkozen.</s>
P: Por que você traduz primeiro o conjunto de dados OASST1/2 completo? Não seria mais rápido traduzir apenas os tópicos com classificação mais alta?
R: Embora você possa ganhar bastante em termos de tempo de transferência primeiro criando os threads e depois traduzindo-os, fornecemos traduções completas do OASST1/2 para a comunidade, pois acreditamos que elas podem ser úteis por si mesmas.
P: Qual é o desempenho dos ajustes finos em comparação com o vanilla LLaMa3?
R: Embora não tenhamos referências formais, fazer com que o LLaMa3 fale consistentemente outro idioma que não o inglês é um desafio, se não impossível. A língua não inglesa que produz é muitas vezes quebrada gramaticalmente. Nossos ajustes finos não mostram esse comportamento.
P: Posso usar outras estruturas para ajuste fino?
R: Sim, você pode, usamos Axolotl para treinamento em configurações multi-GPU.
P: Posso misturar diferentes modelos de tradução?
R: Com certeza, achamos que pode até aumentar o desempenho se a tradução for feita por vários modelos. Você pode conseguir isso interrompendo antecipadamente uma tradução e continuando a partir dos pontos de verificação, executando novamente o script de tradução com um modelo de tradução diferente.
Estamos ativamente à procura de financiamento para democratizar a IA e fazer avançar as suas aplicações. Contate-nos em [email protected] se quiser investir.
