Home

Awesome

RGB

News

Quick links

Environment

conda create -n rgb python=3.10.0
conda activate rgb
bash env.sh

Retrieval-Augmented Generation Benchmark

The data is putted in data/

data/
├── en.json
├── en_refine.json
├── en_int.json
├── en_fact.json
├── zh.json
├── zh_refine.json
├── zh_int.json
└── zh_fact.json

To evalute the Information Integration, you should use zh_int or en_int for Chinese questions or English questions.

To evalute the Counterfactual Robustness, you should use zh_fact or en_fact for Chinese questions or English questions.

The refined data

We refine the retrieved documents and some answers of en.json and zh.json, and name the new data files as en_refine.json and zh_refine.json:

Evaluation

For evaluating ChatGPT, you can run as:

python evalue.py \
--dataset en \
--modelname chatgpt \
--temp 0.2 \
--noise_rate 0.6 \
--api_key YourAPIKEY \
--passage_num 5

For evaluating other models, you can run as:

python evalue.py \
--dataset en \
--modelname chatglm2-6b \
--temp 0.2 \
--noise_rate 0.6 \
--plm THUDM/chatglm-6b \
--passage_num 5

You should change modelname and plm for different models, where plm is the path of model.

temp is the temperature of model.

noise_rate is rate of noisy documents in inputs.

passage_num is number of provided documents for LLM (default is 5).

The outputs are:


To evaluate rejection using ChatGPT, you should first run the evalue.py in noise_rate=1 to obtain the generation result, and then run:

python reject_evalue.py \
--dataset en \
--modelname chatglm2-6b \
--api_key YourAPIKEY

The "reject_rate" in the outputs are the reject rate (Rej*).


To evaluate counterfactual robustness using ChatGPT, you should first run the evalue.py in dataset=en_fact/zh_fact to obtain the generation result, and then run:

python fact_evalue.py \
--dataset en_fact \
--modelname chatglm2-6b \
--api_key YourAPIKEY

The "reject_rate" in the outputs are the error detection rates (ED*). The correct_rate in the outputs are the error correction rate (CR)

License

The code and data are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International Public License for Noncommercial use only. Any commercial use should get formal permission first.

Shield: CC BY-NC-SA 4.0

This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.

CC BY-NC-SA 4.0