Adapting LLMs to Domains via Continual Pre-Training (ICLR 2024)

This repo contains the domain-specific base model developed from LLaMA-1-13B, using the method in our 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.

[2024/11/29] 🤗 Introduce the multimodal version of AdaptLLM at AdaMLLM, for adapting MLLMs to domains 🤗

**************************** Updates ****************************

2024/11/29: Released AdaMLLM for adapting MLLMs to domains
2024/9/20: Our research paper for Instruction-Pretrain has been accepted by EMNLP 2024
2024/8/29: Updated guidelines on evaluating any 🤗Huggingface models on the domain-specific tasks
2024/6/22: Released the benchmarking code
2024/6/21: Released the general version of AdaptLLM at Instruction-Pretrain
2024/4/2: Released the raw data splits (train and test) of all the evaluation datasets
2024/1/16: Our research paper for AdaptLLM 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

1. Domain-Specific Models

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.

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 model (💗 An amazing usage example):

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("AdaptLLM/medicine-LLM-13B")
tokenizer = AutoTokenizer.from_pretrained("AdaptLLM/medicine-LLM-13B", use_fast=False)

# 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.'''

# Simply use your input as the prompt for base models
prompt = user_input

inputs = tokenizer(prompt, return_tensors="pt", add_special_tokens=False).input_ids.to(model.device)
outputs = model.generate(input_ids=inputs, max_length=2048)[0]

answer_start = int(inputs.shape[-1])
pred = tokenizer.decode(outputs[answer_start:], skip_special_tokens=True)

print(pred)

LLaMA-3-8B (💡New!)

In our recent research on Instruction-Pretrain, we developed a context-based instruction synthesizer to augment the raw corpora with instruction-response pairs, enabling Llama3-8B to be comparable to or even outperform Llama3-70B: Finance-Llama3-8B, Biomedicine-Llama3-8B.

2. Domain-Specific Tasks

Pre-templatized Testing Splits

To easily reproduce our prompting results, we have uploaded the filled-in zero/few-shot input instructions and output completions of the test 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.

Evaluating Any Huggingface LMs on Domain-Specific Tasks (💡New!)

You can use the following script to reproduce our results and evaluate any other Huggingface models on domain-specific tasks. Note that the script is NOT applicable to models that require specific prompt templates (e.g., Llama2-chat, Llama3-Instruct).

1). Set Up Dependencies

git clone https://github.com/microsoft/LMOps
cd LMOps/adaptllm
pip install -r requirements.txt

2). Evaluate the Model

# Select the domain from ['biomedicine', 'finance', 'law']
DOMAIN='biomedicine'
  
# Specify any Huggingface model name (Not applicable to chat models)
MODEL='AdaptLLM/medicine-LLM-13B'
  
# Model parallelization:
# - Set MODEL_PARALLEL=False if the model fits on a single GPU. 
#   We observe that LMs smaller than 10B always meet this requirement.
# - Set MODEL_PARALLEL=True if the model is too large and encounters OOM on a single GPU.
MODEL_PARALLEL=True
  
# Choose the number of GPUs from [1, 2, 4, 8]
N_GPU=2
  
# Whether to add a BOS token at the beginning of the prompt input:
# - Set to False for AdaptLLM.
# - Set to True for instruction-pretrain models.
# If unsure, we recommend setting it to False, as this is suitable for most LMs.
add_bos_token=False

# Run the evaluation script
bash scripts/inference.sh ${DOMAIN} ${MODEL} ${add_bos_token} ${MODEL_PARALLEL} ${N_GPU}

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}
}

AdaptLLM
/

medicine-LLM-13B