Datasets:
Update README.md
Browse files
README.md
CHANGED
@@ -30,24 +30,25 @@ tags:
|
|
30 |
- finance
|
31 |
---
|
32 |
|
33 |
-
# Adapting
|
34 |
-
This repo contains the **evaluation datasets** for our
|
35 |
|
36 |
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**.
|
37 |
|
38 |
-
###
|
39 |
|
40 |
**************************** **Updates** ****************************
|
41 |
-
* 2024/
|
42 |
-
* 2024/6/
|
43 |
-
* 2024/
|
44 |
-
* 2024/
|
45 |
-
*
|
46 |
-
* 2023/12/
|
47 |
-
* 2023/
|
|
|
48 |
|
49 |
|
50 |
-
## Domain-Specific
|
51 |
### LLaMA-1-7B
|
52 |
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:
|
53 |
|
@@ -58,78 +59,58 @@ In our paper, we develop three domain-specific models from LLaMA-1-7B, which are
|
|
58 |
### LLaMA-1-13B
|
59 |
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).
|
60 |
|
61 |
-
|
62 |
-
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)
|
63 |
|
64 |
-
|
65 |
-
|
66 |
-
from transformers import AutoModelForCausalLM, AutoTokenizer
|
67 |
|
68 |
-
|
69 |
-
tokenizer = AutoTokenizer.from_pretrained("AdaptLLM/finance-chat")
|
70 |
|
71 |
-
|
72 |
-
user_input = '''Use this fact to answer the question: Title of each class Trading Symbol(s) Name of each exchange on which registered
|
73 |
-
Common Stock, Par Value $.01 Per Share MMM New York Stock Exchange
|
74 |
-
MMM Chicago Stock Exchange, Inc.
|
75 |
-
1.500% Notes due 2026 MMM26 New York Stock Exchange
|
76 |
-
1.750% Notes due 2030 MMM30 New York Stock Exchange
|
77 |
-
1.500% Notes due 2031 MMM31 New York Stock Exchange
|
78 |
-
|
79 |
-
Which debt securities are registered to trade on a national securities exchange under 3M's name as of Q2 of 2023?'''
|
80 |
-
|
81 |
-
# Apply the prompt template and system prompt of LLaMA-2-Chat demo for chat models (NOTE: NO prompt template is required for base models!)
|
82 |
-
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
|
83 |
-
prompt = f"<s>[INST] <<SYS>>{our_system_prompt}<</SYS>>\n\n{user_input} [/INST]"
|
84 |
-
|
85 |
-
# # NOTE:
|
86 |
-
# # If you want to apply your own system prompt, please integrate it into the instruction part following our system prompt like this:
|
87 |
-
# your_system_prompt = "Please, check if the answer can be inferred from the pieces of context provided."
|
88 |
-
# prompt = f"<s>[INST] <<SYS>>{our_system_prompt}<</SYS>>\n\n{your_system_prompt}\n{user_input} [/INST]"
|
89 |
-
|
90 |
-
inputs = tokenizer(prompt, return_tensors="pt", add_special_tokens=False).input_ids.to(model.device)
|
91 |
-
outputs = model.generate(input_ids=inputs, max_length=4096)[0]
|
92 |
-
|
93 |
-
answer_start = int(inputs.shape[-1])
|
94 |
-
pred = tokenizer.decode(outputs[answer_start:], skip_special_tokens=True)
|
95 |
-
|
96 |
-
print(f'### User Input:\n{user_input}\n\n### Assistant Output:\n{pred}')
|
97 |
-
```
|
98 |
-
|
99 |
-
## Domain-Specific Tasks
|
100 |
-
|
101 |
-
### Pre-templatized/Formatted Testing Splits
|
102 |
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).
|
103 |
|
104 |
**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.
|
105 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
106 |
### Raw Datasets
|
107 |
-
We have also uploaded the raw training and testing splits, for facilitating fine-tuning or other usages
|
108 |
-
- [ChemProt](https://huggingface.co/datasets/AdaptLLM/ChemProt)
|
109 |
-
- [RCT](https://huggingface.co/datasets/AdaptLLM/RCT)
|
110 |
-
- [ConvFinQA](https://huggingface.co/datasets/AdaptLLM/ConvFinQA)
|
111 |
-
- [FiQA_SA](https://huggingface.co/datasets/AdaptLLM/FiQA_SA)
|
112 |
-
- [Headline](https://huggingface.co/datasets/AdaptLLM/Headline)
|
113 |
-
- [NER](https://huggingface.co/datasets/AdaptLLM/NER)
|
114 |
-
- [FPB](https://huggingface.co/datasets/AdaptLLM/FPB)
|
115 |
-
|
116 |
-
The other datasets used in our paper have already been available in huggingface, and you can directly load them with the following code:
|
117 |
-
```python
|
118 |
-
from datasets import load_dataset
|
119 |
-
|
120 |
-
# MQP:
|
121 |
-
dataset = load_dataset('medical_questions_pairs')
|
122 |
-
# PubmedQA:
|
123 |
-
dataset = load_dataset('bigbio/pubmed_qa')
|
124 |
-
# USMLE:
|
125 |
-
dataset=load_dataset('GBaker/MedQA-USMLE-4-options')
|
126 |
-
# SCOTUS
|
127 |
-
dataset = load_dataset("lex_glue", 'scotus')
|
128 |
-
# CaseHOLD
|
129 |
-
dataset = load_dataset("lex_glue", 'case_hold')
|
130 |
-
# UNFAIR-ToS
|
131 |
-
dataset = load_dataset("lex_glue", 'unfair_tos')
|
132 |
-
```
|
133 |
|
134 |
## Citation
|
135 |
If you find our work helpful, please cite us:
|
|
|
30 |
- finance
|
31 |
---
|
32 |
|
33 |
+
# Adapting LLM to Domains (ICLR 2024)
|
34 |
+
This repo contains the **evaluation datasets** for our paper [Adapting Large Language Models via Reading Comprehension](https://huggingface.co/papers/2309.09530).
|
35 |
|
36 |
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**.
|
37 |
|
38 |
+
### [2024/6/21] 🤗 We release the 2nd version of AdaptLLM at [Instruction-Pretrain](https://huggingface.co/instruction-pretrain), effective for both pre-training from scratch and continual pre-training 🤗
|
39 |
|
40 |
**************************** **Updates** ****************************
|
41 |
+
* 2024/8/29: Updated [guidelines](https://huggingface.co/datasets/AdaptLLM/finance-tasks) on evaluating any 🤗Huggingface models on the domain-specific tasks
|
42 |
+
* 2024/6/22: Released the [benchmarking code](https://github.com/microsoft/LMOps/tree/main/adaptllm)
|
43 |
+
* 2024/6/21: Released the 2nd version of AdaptLLM at [Instruction-Pretrain](https://huggingface.co/instruction-pretrain)
|
44 |
+
* 2024/4/2: Released the [raw data splits (train and test)](https://huggingface.co/datasets/AdaptLLM/ConvFinQA) of all the evaluation datasets
|
45 |
+
* 2024/1/16: Our [research paper](https://huggingface.co/papers/2309.09530) has been accepted by ICLR 2024
|
46 |
+
* 2023/12/19: Released our [13B base models](https://huggingface.co/AdaptLLM/law-LLM-13B) developed from LLaMA-1-13B
|
47 |
+
* 2023/12/8: Released our [chat models](https://huggingface.co/AdaptLLM/law-chat) developed from LLaMA-2-Chat-7B
|
48 |
+
* 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
|
49 |
|
50 |
|
51 |
+
## 1. Domain-Specific Models
|
52 |
### LLaMA-1-7B
|
53 |
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:
|
54 |
|
|
|
59 |
### LLaMA-1-13B
|
60 |
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).
|
61 |
|
62 |
+
### LLaMA-2-Chat
|
63 |
+
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).
|
64 |
|
65 |
+
### LLaMA-3-8B (💡New!)
|
66 |
+
In our recent research on [Instruction-Pretrain](https://huggingface.co/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](https://huggingface.co/instruction-pretrain/finance-Llama3-8B), [Biomedicine-Llama3-8B](https://huggingface.co/instruction-pretrain/medicine-Llama3-8B).
|
|
|
67 |
|
68 |
+
## 2. Domain-Specific Tasks
|
|
|
69 |
|
70 |
+
### Pre-templatized Testing Splits
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
71 |
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).
|
72 |
|
73 |
**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.
|
74 |
|
75 |
+
### Evaluating Any Huggingface LMs on Domain-Specific Tasks (💡New!)
|
76 |
+
You can use the following scripts to reproduce our results and evaluate any other Huggingface models on the testing splits:
|
77 |
+
|
78 |
+
1). **Set Up Dependencies**
|
79 |
+
```bash
|
80 |
+
git clone https://github.com/microsoft/LMOps
|
81 |
+
cd LMOps/adaptllm
|
82 |
+
pip install -r requirements.txt
|
83 |
+
```
|
84 |
+
|
85 |
+
2). **Evaluate the Model**
|
86 |
+
```bash
|
87 |
+
# Select the domain from ['biomedicine', 'finance', 'law']
|
88 |
+
DOMAIN='finance'
|
89 |
+
|
90 |
+
# Specify any Huggingface model name (Not applicable to chat models)
|
91 |
+
MODEL='AdaptLLM/finance-LLM'
|
92 |
+
|
93 |
+
# Model parallelization:
|
94 |
+
# - Set MODEL_PARALLEL=False if the model fits on a single GPU.
|
95 |
+
# We observe that LMs smaller than 10B always meet this requirement.
|
96 |
+
# - Set MODEL_PARALLEL=True if the model is too large and encounters OOM on a single GPU.
|
97 |
+
MODEL_PARALLEL=False
|
98 |
+
|
99 |
+
# Choose the number of GPUs from [1, 2, 4, 8]
|
100 |
+
N_GPU=1
|
101 |
+
|
102 |
+
# Whether to add a BOS token at the beginning of the prompt input:
|
103 |
+
# - Set to False for AdaptLLM.
|
104 |
+
# - Set to True for instruction-pretrain models.
|
105 |
+
# If unsure, we recommend setting it to False, as this is suitable for most LMs.
|
106 |
+
add_bos_token=False
|
107 |
+
|
108 |
+
# Run the evaluation script
|
109 |
+
bash scripts/inference.sh ${DOMAIN} ${MODEL} ${add_bos_token} ${MODEL_PARALLEL} ${N_GPU}
|
110 |
+
```
|
111 |
+
|
112 |
### Raw Datasets
|
113 |
+
We have also uploaded the [raw training and testing splits](https://huggingface.co/datasets/AdaptLLM/ConvFinQA), for facilitating fine-tuning or other usages.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
114 |
|
115 |
## Citation
|
116 |
If you find our work helpful, please cite us:
|