Darwin

Other source code

1.0.0

Download

Darwin: A Tailored GPT for the Scientific Domain ??

logo

Organization: University of New South Wales（UNSW) AI4Science & GreenDynamics AI

Darwin is an open-source project dedicated to pretrain and fine-tune the LLaMA model on scientific literature and datasets. Specifically designed for the scientific domain with an emphasis on materials science, chemistry, and physics, Darwin integrates structured and unstructured scientific knowledge to enhance the efficacy of language models in scientific research.

Usage and License Notices: Darwin is licensed and intended for research use only. The dataset is licensed under CC BY NC 4.0, allowing non-commercial use. Models trained using this dataset should not be used outside of research purposes. The weight diff is also under CC BY NC 4.0 license

Update

[2024.11.20]

Key Achievements

Proved that Darwin’s fine-tuning strategies (QA + multi-task) substantially improve performance on diverse ML tasks.
Established Darwin as a competitive model, bridging the gap between specialized ML methods and large-scale generalist models like GPT-4.

Model Performance Insights

Comparison of QA + Multi-task Strategies Across LLaMA Variants
- Conducted extensive comparisons of QA and multi-task fine-tuning strategies on LLaMA1, LLaMA2, LLaMA3, and LLaMA3.1 models.
- Finding: LLaMA1 with QA + multi-task fine-tuning achieves the best performance, outperforming all other variants.
Evaluation Against Other Models
- Demonstrated that Darwin consistently surpasses most ML methods, GPT-3.5 fine-tuned models, and even GPT-4 in few-shot learning tasks.
- Although some specialized models still maintain state-of-the-art results, Darwin achieves competitive performance across a broad range of tasks.
Comparison of Full Fine-tuning vs. LoRA
- Investigated the use of LoRA fine-tuning and observed significantly lower performance compared to full fine-tuning.
SFT on Non-pretrained Architectures
- Successfully applied supervised fine-tuning (SFT) on non-pretrained LLaMA architectures, proving that models can acquire domain-specific knowledge effectively through fine-tuning alone.

Data Strategies and Insights

Impact of QA Data on Model Performance
- Verified that both QA fine-tuning and multi-task learning improve model performance, not only for LLaMA but also for other architectures like Mistral.
- Mixing QA data with general data improves model performance without causing model annealing.
Synchronized Data and Format Matching
- Tested the use of synchronized (sync) data for target tasks.
- Findings:
  - Sync data with similar format improves performance significantly.
  - Sync data with differing formats degrades performance.
Multi-task Mixing Strategies (?)
- Explored various multi-task data mixing strategies and found that ranking data by difficulty (using single-task loss as a metric) outperforms random data mixing strategies.

[2024.02.15] SOTA in MatBench by Material Projects: DARWIN is the SOTA model in experimental bandgap prediction tasks and metallic classification tasks, better than Fine-tuned GPT3.5 and dedicated ML models. https://matbench.materialsproject.org/Leaderboards%20Per-Task/matbench_v0.1_matbench_expt_gap/

☆ [2023.09.15]Google Colab Version available: Try our DARWIN with Google Colab: inference.ipynb

Model Overview

Darwin, based on the 7B LLaMA model, is trained on over 100,000 instruction-following data points generated by the Darwin Scientific Instruction Generator (SIG) from various scientific FAIR datasets and a literature corpus. By focusing on the factual correctness of the model's responses, Darwin represents a significant stride towards leveraging Large Language Models (LLMs) for scientific discovery. Preliminary human evaluations indicate that Darwin 7B outperforms GPT-4 in scientific Q&A and fine-tuned GPT-3 in solving chemistry problems (like gptChem).

We are actively developing Darwin for more advanced scientific domain experiments, and we're also integrating Darwin with LangChain to solve more complex scientific tasks (like a private research assistant for personal computers).

Please note, Darwin is still under development, and many limitations need to be addressed. Most importantly, we have yet to fine-tune Darwin for maximum safety. We encourage users to report any concerning behavior to help improve the model's safety and ethical considerations.

DARWIN.mp4

DEMO LINK

Model Comparison

compare

Getting Started

Installation

First install the requirements:

pip install -r requirements.txt

Preparing the Darwin Weights

Download the checkpoints of the Darwin-7B Weights from onedrive. Once you've downloaded the model, you can try our demo:

python inference.py <your path to darwin-7b>

Please note, the inference requires at least 10GB of GPU memory for Darwin 7B.

Fine-tuning

To further fine-tune our Darwin-7b with different datasets, below is a command that works on a machine with 4 A100 80G GPUs.

torchrun  --nproc_per_node=8 --master_port=1212 train.py 
    --model_name_or_path <your path to darwin-7b> 
    --data_path <your path to dataset> 
    --bf16 True 
    --output_dir <your output dir> 
    --num_train_epochs 3 
    --per_device_train_batch_size 1 
    --per_device_eval_batch_size 1 
    --gradient_accumulation_steps 1 
    --evaluation_strategy "no" 
    --save_strategy "steps" 
    --save_steps 500 
    --save_total_limit 1 
    --learning_rate 2e-5 
    --weight_decay 0. 
    --warmup_ratio 0.03 
    --lr_scheduler_type "cosine" 
    --logging_steps 1 
    --fsdp "full_shard auto_wrap" 
    --fsdp_transformer_layer_cls_to_wrap 'LlamaDecoderLayer' 
    --tf32 False

Datasets Information

Our data comes from two primary sources:

A raw literature corpus containing 6.0M papers on materials science, chemistry, and physics was published after 2000. The publishers include ACS, RSC, Springer Nature, Wiley, and Elsevier. We thank them for their support.

FAIR Datasets - We've collected data from 16 FAIR Datasets.

Data Generation

We developed Darwin-SIG to generate scientific instructions. It can memorize long texts from full literature texts (average ~5000 words) and generate question-and-answer (Q&A) data based on scientific literature keywords (from web of science API)

Note: You could also use GPT3.5 or GPT-4 for generation, but these options might be costly.

Please be aware that we can't share the training dataset due to agreements with the publishers.

Authors

This project is a collaborative effort by the following:

UNSW & GreenDynamics: Tong Xie, Shaozhou Wang

UNSW: Imran Razzak, Cody Huang

USYD & DARE Centre: Clara Grazian

GreenDynamics: Yuwei Wan,Yixuan Liu

Bram Hoex and Wenjie Zhang from UNSW Engineering advised all.

Citation

If you use the data or code from this repository in your work, please cite it accordingly.

DAWRIN Foundational Large Language Model & Semi-Self Instruct Fine Tuning

@misc{xie2023darwin,
      title={DARWIN Series: Domain Specific Large Language Models for Natural Science}, 
      author={Tong Xie and Yuwei Wan and Wei Huang and Zhenyu Yin and Yixuan Liu and Shaozhou Wang and Qingyuan Linghu and Chunyu Kit and Clara Grazian and Wenjie Zhang and Imran Razzak and Bram Hoex},
      year={2023},
      eprint={2308.13565},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Fine-tuned GPT-3 & LLaMA for Material Discovery (Single Task Training)

@article{xie2023large,
  title={Large Language Models as Master Key: Unlocking the Secrets of Materials Science},
  author={Xie, Tong and Wan, Yuwei and Zhou, Yufei and Huang, Wei and Liu, Yixuan and Linghu, Qingyuan and Wang, Shaozhou and Kit, Chunyu and Grazian, Clara and Zhang, Wenjie and others},
  journal={Available at SSRN 4534137},
  year={2023}
}