--- configs: - config_name: SCOTUS data_files: - split: test path: scotus/test.json - config_name: CaseHOLD data_files: - split: test path: case_hold/test.json - config_name: UNFAIR_ToS data_files: - split: test path: unfair_tos/test.json task_categories: - text-classification - question-answering - zero-shot-classification language: - en tags: - legal --- # Domain Adaptation of Large Language Models This repo contains the **evaluation datasets** for our **ICLR 2024** paper [Adapting Large Language Models via Reading Comprehension](https://huggingface.co/papers/2309.09530). 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**. ### 🤗 We are currently working hard on developing models across different domains, scales and architectures! Please stay tuned! 🤗 **************************** **Updates** **************************** * 2024/4/2: Released the raw data splits (train and test) of all the evaluation datasets * 2024/1/16: 🎉 Our [research paper](https://huggingface.co/papers/2309.09530) has been accepted by ICLR 2024!!!🎉 * 2023/12/19: Released our [13B base models](https://huggingface.co/AdaptLLM/law-LLM-13B) developed from LLaMA-1-13B. * 2023/12/8: Released our [chat models](https://huggingface.co/AdaptLLM/law-chat) developed from LLaMA-2-Chat-7B. * 2023/9/18: Released our [paper](https://huggingface.co/papers/2309.09530), [code](https://github.com/microsoft/LMOps), [data](https://huggingface.co/datasets/AdaptLLM/law-tasks), and [base models](https://huggingface.co/AdaptLLM/law-LLM) 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](https://huggingface.co/AdaptLLM/medicine-LLM), [Finance-LLM](https://huggingface.co/AdaptLLM/finance-LLM) and [Law-LLM](https://huggingface.co/AdaptLLM/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](https://huggingface.co/AdaptLLM/medicine-LLM-13B), [Finance-LLM-13B](https://huggingface.co/AdaptLLM/finance-LLM-13B) and [Law-LLM-13B](https://huggingface.co/AdaptLLM/law-LLM-13B). ## Domain-Specific LLaMA-2-Chat Our method is also effective for aligned models! LLaMA-2-Chat requires a [specific data format](https://huggingface.co/blog/llama2#how-to-prompt-llama-2), 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](https://huggingface.co/AdaptLLM/medicine-chat), [Finance-Chat](https://huggingface.co/AdaptLLM/finance-chat) and [Law-Chat](https://huggingface.co/AdaptLLM/law-chat) For example, to chat with the law-chat model: ```python from transformers import AutoModelForCausalLM, AutoTokenizer model = AutoModelForCausalLM.from_pretrained("AdaptLLM/law-chat") tokenizer = AutoTokenizer.from_pretrained("AdaptLLM/law-chat") # Put your input here: user_input = '''Question: Which of the following is false about ex post facto laws? Options: - They make criminal an act that was innocent when committed. - They prescribe greater punishment for an act than was prescribed when it was done. - They increase the evidence required to convict a person than when the act was done. - They alter criminal offenses or punishment in a substantially prejudicial manner for the purpose of punishing a person for some past activity. 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"[INST] <>{our_system_prompt}<>\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"[INST] <>{our_system_prompt}<>\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 ### Pre-templatized/Formatted 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](https://huggingface.co/datasets/AdaptLLM/medicine-tasks), [finance-tasks](https://huggingface.co/datasets/AdaptLLM/finance-tasks), and [law-tasks](https://huggingface.co/datasets/AdaptLLM/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. ### Raw Datasets We have also uploaded the raw training and testing splits, for facilitating fine-tuning or other usages: - [ChemProt](https://huggingface.co/datasets/AdaptLLM/ChemProt) - [RCT](https://huggingface.co/datasets/AdaptLLM/RCT) - [ConvFinQA](https://huggingface.co/datasets/AdaptLLM/ConvFinQA) - [FiQA_SA](https://huggingface.co/datasets/AdaptLLM/FiQA_SA) - [Headline](https://huggingface.co/datasets/AdaptLLM/Headline) - [NER](https://huggingface.co/datasets/AdaptLLM/NER) The other datasets used in our paper have already been available in huggingface, so you can directly load them with the following code ```python from datasets import load_dataset # MQP: dataset = load_dataset('medical_questions_pairs') # PubmedQA: dataset = load_dataset('bigbio/pubmed_qa') # SCOTUS dataset = load_dataset("lex_glue", 'scotus') # CaseHOLD dataset = load_dataset("lex_glue", 'case_hold') # UNFAIR-ToS dataset = load_dataset("lex_glue", 'unfair_tos') ``` ## Citation If you find our work helpful, please cite us: ```bibtex @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} } ```