Adapting Large Language Models to Domains via Continual Pre-Training
This repo contains the domain-specific chat model developed from LLaMA-2-Chat-7B, using the method in our ICLR 2024 paper Adapting Large Language Models via Reading Comprehension.
We explore continued pre-training on domain-specific corpora for large language models. While this approach enriches LLMs with domain knowledge, it significantly hurts their prompting ability for question answering. Inspired by human learning via reading comprehension, we propose a simple method to transform large-scale pre-training corpora into reading comprehension texts, consistently improving prompting performance across tasks in biomedicine, finance, and law domains. Our 7B model competes with much larger domain-specific models like BloombergGPT-50B.
🤗 [2024/6/21] We release the 2nd version of AdaptLLM at Instruction-Pretrain, effective for both pre-training from scratch and continual pre-training 🤗
**************************** Updates ****************************
- 2024/6/22: Released the benchmarking code.
- 2024/6/21: 👏🏻 Released the 2nd version of AdaptLLM at Instruction-Pretrain 👏🏻
- 2024/1/16: 🎉 Our research paper has been accepted by ICLR 2024 🎉
- 2023/12/19: Released our 13B base models developed from LLaMA-1-13B.
- 2023/12/8: Released our chat models developed from LLaMA-2-Chat-7B.
- 2023/9/18: Released our paper, code, data, and base models developed from LLaMA-1-7B.
Domain-Specific LLaMA-1
LLaMA-1-7B
In our paper, we develop three domain-specific models from LLaMA-1-7B, which are also available in Huggingface: Biomedicine-LLM, Finance-LLM and Law-LLM, the performances of our AdaptLLM compared to other domain-specific LLMs are:
LLaMA-1-13B
Moreover, we scale up our base model to LLaMA-1-13B to see if our method is similarly effective for larger-scale models, and the results are consistently positive too: Biomedicine-LLM-13B, Finance-LLM-13B and Law-LLM-13B.
Domain-Specific LLaMA-2-Chat
Our method is also effective for aligned models! LLaMA-2-Chat requires a specific data format, and our reading comprehension can perfectly fit the data format by transforming the reading comprehension into a multi-turn conversation. We have also open-sourced chat models in different domains: Biomedicine-Chat, Finance-Chat and Law-Chat
For example, to chat with the biomedicine-chat model:
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("AdaptLLM/medicine-chat")
tokenizer = AutoTokenizer.from_pretrained("AdaptLLM/medicine-chat")
# Put your input here:
user_input = '''Question: Which of the following is an example of monosomy?
Options:
- 46,XX
- 47,XXX
- 69,XYY
- 45,X
Please provide your choice first and then provide explanations if possible.'''
# Apply the prompt template and system prompt of LLaMA-2-Chat demo for chat models (NOTE: NO prompt template is required for base models!)
our_system_prompt = "\nYou are a helpful, respectful and honest assistant. Always answer as helpfully as possible, while being safe. Your answers should not include any harmful, unethical, racist, sexist, toxic, dangerous, or illegal content. Please ensure that your responses are socially unbiased and positive in nature.\n\nIf a question does not make any sense, or is not factually coherent, explain why instead of answering something not correct. If you don't know the answer to a question, please don't share false information.\n" # Please do NOT change this
prompt = f"<s>[INST] <<SYS>>{our_system_prompt}<</SYS>>\n\n{user_input} [/INST]"
# # NOTE:
# # If you want to apply your own system prompt, please integrate it into the instruction part following our system prompt like this:
# your_system_prompt = "Please, answer this question faithfully."
# prompt = f"<s>[INST] <<SYS>>{our_system_prompt}<</SYS>>\n\n{your_system_prompt}\n{user_input} [/INST]"
inputs = tokenizer(prompt, return_tensors="pt", add_special_tokens=False).input_ids.to(model.device)
outputs = model.generate(input_ids=inputs, max_length=4096)[0]
answer_start = int(inputs.shape[-1])
pred = tokenizer.decode(outputs[answer_start:], skip_special_tokens=True)
print(f'### User Input:\n{user_input}\n\n### Assistant Output:\n{pred}')
Domain-Specific Tasks
To easily reproduce our results, we have uploaded the filled-in zero/few-shot input instructions and output completions of each domain-specific task: biomedicine-tasks, finance-tasks, and law-tasks.
Note: those filled-in instructions are specifically tailored for models before alignment and do NOT fit for the specific data format required for chat models.
Open LLM Leaderboard Evaluation Results
Detailed results can be found here
Metric | Value |
---|---|
Avg. | 52.99 |
AI2 Reasoning Challenge (25-Shot) | 53.75 |
HellaSwag (10-Shot) | 76.11 |
MMLU (5-Shot) | 49.98 |
TruthfulQA (0-shot) | 43.46 |
Winogrande (5-shot) | 75.69 |
GSM8k (5-shot) | 18.95 |
Citation
If you find our work helpful, please cite us:
@inproceedings{
cheng2024adapting,
title={Adapting Large Language Models via Reading Comprehension},
author={Daixuan Cheng and Shaohan Huang and Furu Wei},
booktitle={The Twelfth International Conference on Learning Representations},
year={2024},
url={https://openreview.net/forum?id=y886UXPEZ0}
}
- Downloads last month
- 3,854
Model tree for AdaptLLM/medicine-chat
Datasets used to train AdaptLLM/medicine-chat
Spaces using AdaptLLM/medicine-chat 2
Evaluation results
- normalized accuracy on AI2 Reasoning Challenge (25-Shot)test set Open LLM Leaderboard53.750
- normalized accuracy on HellaSwag (10-Shot)validation set Open LLM Leaderboard76.110
- accuracy on MMLU (5-Shot)test set Open LLM Leaderboard49.980
- mc2 on TruthfulQA (0-shot)validation set Open LLM Leaderboard43.460
- accuracy on Winogrande (5-shot)validation set Open LLM Leaderboard75.690
- accuracy on GSM8k (5-shot)test set Open LLM Leaderboard18.950