Update README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,123 @@
|
|
1 |
-
---
|
2 |
-
license: mit
|
3 |
-
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: mit
|
3 |
+
---
|
4 |
+
# ExtAgents
|
5 |
+
|
6 |
+
<p align="center">
|
7 |
+
<a href="https://github.com/THUNLP-MT/ExtAgents">π Github</a> |
|
8 |
+
<a href="https://arxiv.org/abs/2505.21471">π Paper</a> |
|
9 |
+
<a href="https://huggingface.co/datasets/zhennan1/ExtAgents">π€ Data</a>
|
10 |
+
</p>
|
11 |
+
|
12 |
+
## Introduction
|
13 |
+
|
14 |
+
ExtAgents is a framework for scaling external knowledge input beyond the context length of LLMs via multi-agent collaboration.
|
15 |
+
|
16 |
+
## Setup
|
17 |
+
|
18 |
+
```bash
|
19 |
+
conda create -n extagents python=3.10 -y
|
20 |
+
conda activate extagents
|
21 |
+
pip install -r requirements.txt
|
22 |
+
```
|
23 |
+
|
24 |
+
## Data
|
25 |
+
|
26 |
+
You can download the data with the script:
|
27 |
+
|
28 |
+
```bash
|
29 |
+
bash scripts/download_data.sh
|
30 |
+
```
|
31 |
+
|
32 |
+
Or you can download the data manually from one of the following links:
|
33 |
+
|
34 |
+
- [Google Drive](https://drive.google.com/drive/folders/1FQSojqgF1VdumXxSh1UbIoE6lQ2E_xJn?usp=sharing)
|
35 |
+
- [Tsinghua Cloud](https://cloud.tsinghua.edu.cn/d/b8aab568cf5c4785b457/)
|
36 |
+
|
37 |
+
The data should be organized as follows:
|
38 |
+
|
39 |
+
```bash
|
40 |
+
./
|
41 |
+
βββ data/
|
42 |
+
βββ sampled_hotpot_questions.json
|
43 |
+
βββ rag_1000k.jsonl
|
44 |
+
βββ longbook_qa_eng.jsonl
|
45 |
+
βββ longbook_qa_chn.jsonl
|
46 |
+
```
|
47 |
+
|
48 |
+
## Usage
|
49 |
+
|
50 |
+
### Generation
|
51 |
+
|
52 |
+
We currently support three tasks: RAG, En.QA, Zh.QA.
|
53 |
+
|
54 |
+
The RAG task is a question answering task, where the input is a question and a context. The question and answer are sampled from the [HotpotQA](https://github.com/hotpotqa/hotpot). The context is a long text, which is the concatenation of documents retrieved from Wikipedia using BM25 embedding. We use [KILT knowledge source](http://dl.fbaipublicfiles.com/KILT/kilt_knowledgesource.json) as our knowledge source. It is based on the [2019/08/01 Wikipedia dump](http://dl.fbaipublicfiles.com/BLINK/enwiki-pages-articles.xml.bz2). We have provided the context in the `data/rag_1000k.jsonl` file.
|
55 |
+
|
56 |
+
The En.QA and Zh.QA tasks are question answering tasks, where the input is a question and a context. The question, answer and context are from the [InfiniteBench](https://github.com/OpenBMB/InfiniteBench).
|
57 |
+
|
58 |
+
Here is an example command to generate predictions for RAG task:
|
59 |
+
|
60 |
+
```bash
|
61 |
+
python main.py \
|
62 |
+
--task rag \
|
63 |
+
--output_dir results_rag \
|
64 |
+
--chunk_length 8000 \
|
65 |
+
--input_length 128000 \
|
66 |
+
--api_url "YOUR_API_URL" \
|
67 |
+
--api_key "YOUR_API_KEY" \
|
68 |
+
--model "gpt-4o-mini-2024-07-18" \
|
69 |
+
--num_workers 8 \
|
70 |
+
> rag.log
|
71 |
+
```
|
72 |
+
|
73 |
+
The generated predictions will be saved in the `results_rag` directory.
|
74 |
+
|
75 |
+
- `--task`: Task, can be `rag`, `en`, `zh`.
|
76 |
+
- `--output_dir`: Directory to save the generated predictions.
|
77 |
+
- `--chunk_length`: Chunk length.
|
78 |
+
- `--input_length`: Input length.
|
79 |
+
- `--model`: Model to use, default is `gpt-4o-mini-2024-07-18`.
|
80 |
+
- `--api_url`: Your API URL, default is os.getenv("OPENAI_BASE_URL").
|
81 |
+
- `--api_key`: Your API Key, default is os.getenv("OPENAI_API_KEY").
|
82 |
+
- `--num_workers`: Number of workers, each worker will process one example.
|
83 |
+
|
84 |
+
You can also set the environment variables `OPENAI_BASE_URL` and `OPENAI_API_KEY` to avoid typing them in the command line.
|
85 |
+
|
86 |
+
```bash
|
87 |
+
export OPENAI_BASE_URL="YOUR_API_URL"
|
88 |
+
export OPENAI_API_KEY="YOUR_API_KEY"
|
89 |
+
```
|
90 |
+
|
91 |
+
### Evaluation
|
92 |
+
|
93 |
+
We provide a script to evaluate the generated predictions. For RAG task, the evaluation is based on the [HotpotQA](https://github.com/hotpotqa/hotpot). For En.QA and Zh.QA task, the evaluation is based on the [InfiniteBench](https://github.com/OpenBMB/InfiniteBench).
|
94 |
+
|
95 |
+
For RAG task:
|
96 |
+
|
97 |
+
```bash
|
98 |
+
bash scripts/eval_rag.sh /path/to/your/output_dir
|
99 |
+
```
|
100 |
+
|
101 |
+
For En.QA task:
|
102 |
+
|
103 |
+
```bash
|
104 |
+
bash scripts/eval_en.sh /path/to/your/output_dir
|
105 |
+
```
|
106 |
+
|
107 |
+
For Zh.QA task:
|
108 |
+
|
109 |
+
```bash
|
110 |
+
bash scripts/eval_zh.sh /path/to/your/output_dir
|
111 |
+
```
|
112 |
+
|
113 |
+
## Citation
|
114 |
+
|
115 |
+
If you find this project helpful, please cite it as follows:
|
116 |
+
|
117 |
+
```bibtex
|
118 |
+
@article{liu2025extagents,
|
119 |
+
title={Scaling External Knowledge Input Beyond The Context Length of LLMs via Multi-Agent Collaboration},
|
120 |
+
author={Zijun Liu and Zhennan Wan and Peng Li and Ming Yan and Ji Zhang and Fei Huang and Yang Liu},
|
121 |
+
year={2025}
|
122 |
+
}
|
123 |
+
```
|