12b84aeb3b
* update gemini api and add gemini models * add openai models * update CHARM evaluation * add CHARM memorization tasks * add CharmMemSummarizer (output eval details for memorization-independent reasoning analysis * update CHARM readme --------- Co-authored-by: wujiang <wujiang@pjlab.org.cn>
8.7 KiB
CHARM✨ Benchmarking Chinese Commonsense Reasoning of LLMs: From Chinese-Specifics to Reasoning-Memorization Correlations [ACL2024]
Dataset Description
CHARM is the first benchmark for comprehensively and in-depth evaluating the commonsense reasoning ability of large language models (LLMs) in Chinese, which covers both globally known and Chinese-specific commonsense. In addition, the CHARM can evaluate the LLMs' memorization-independent reasoning abilities and analyze the typical errors.
Comparison of commonsense reasoning benchmarks
<html lang="en">| Benchmarks | CN-Lang | CSR | CN-specifics | Dual-Domain | Rea-Mem |
|---|---|---|---|---|---|
| Most benchmarks in davis2023benchmarks | ✘ | ✔ | ✘ | ✘ | ✘ |
| XNLI, XCOPA,XStoryCloze | ✔ | ✔ | ✘ | ✘ | ✘ |
| LogiQA, CLUE, CMMLU | ✔ | ✘ | ✔ | ✘ | ✘ |
| CORECODE | ✔ | ✔ | ✘ | ✘ | ✘ |
| CHARM (ours) | ✔ | ✔ | ✔ | ✔ | ✔ |
"CN-Lang" indicates the benchmark is presented in Chinese language. "CSR" means the benchmark is designed to focus on CommonSense Reasoning. "CN-specific" indicates the benchmark includes elements that are unique to Chinese culture, language, regional characteristics, history, etc. "Dual-Domain" indicates the benchmark encompasses both Chinese-specific and global domain tasks, with questions presented in the similar style and format. "Rea-Mem" indicates the benchmark includes closely-interconnected reasoning and memorization tasks.
🛠️ How to Use
Below are the steps for quickly downloading CHARM and using OpenCompass for evaluation.
1. Download CHARM
git clone https://github.com/opendatalab/CHARM ${path_to_CHARM_repo}
cd ${path_to_opencompass}
mkdir data
ln -snf ${path_to_CHARM_repo}/data/CHARM ./data/CHARM
2. Run Inference and Evaluation
cd ${path_to_opencompass}
# modify config file `configs/eval_charm_rea.py`: uncomment or add models you want to evaluate
python run.py configs/eval_charm_rea.py -r --dump-eval-details
# modify config file `configs/eval_charm_mem.py`: uncomment or add models you want to evaluate
python run.py configs/eval_charm_mem.py -r --dump-eval-details
The inference and evaluation results would be in ${path_to_opencompass}/outputs, like this:
outputs
├── CHARM_mem
│ └── chat
│ └── 20240605_151442
│ ├── predictions
│ │ ├── internlm2-chat-1.8b-turbomind
│ │ ├── llama-3-8b-instruct-lmdeploy
│ │ └── qwen1.5-1.8b-chat-hf
│ ├── results
│ │ ├── internlm2-chat-1.8b-turbomind_judged-by--GPT-3.5-turbo-0125
│ │ ├── llama-3-8b-instruct-lmdeploy_judged-by--GPT-3.5-turbo-0125
│ │ └── qwen1.5-1.8b-chat-hf_judged-by--GPT-3.5-turbo-0125
│ └── summary
│ └── 20240605_205020 # MEMORY_SUMMARY_DIR
│ ├── judged-by--GPT-3.5-turbo-0125-charm-memory-Chinese_Anachronisms_Judgment
│ ├── judged-by--GPT-3.5-turbo-0125-charm-memory-Chinese_Movie_and_Music_Recommendation
│ ├── judged-by--GPT-3.5-turbo-0125-charm-memory-Chinese_Sport_Understanding
│ ├── judged-by--GPT-3.5-turbo-0125-charm-memory-Chinese_Time_Understanding
│ └── judged-by--GPT-3.5-turbo-0125.csv # MEMORY_SUMMARY_CSV
└── CHARM_rea
└── chat
└── 20240605_152359
├── predictions
│ ├── internlm2-chat-1.8b-turbomind
│ ├── llama-3-8b-instruct-lmdeploy
│ └── qwen1.5-1.8b-chat-hf
├── results # REASON_RESULTS_DIR
│ ├── internlm2-chat-1.8b-turbomind
│ ├── llama-3-8b-instruct-lmdeploy
│ └── qwen1.5-1.8b-chat-hf
└── summary
├── summary_20240605_205328.csv # REASON_SUMMARY_CSV
└── summary_20240605_205328.txt
3. Generate Analysis Results
cd ${path_to_CHARM_repo}
# generate Table5, Table6, Table9 and Table10 in https://arxiv.org/abs/2403.14112
PYTHONPATH=. python tools/summarize_reasoning.py ${REASON_SUMMARY_CSV}
# generate Figure3 and Figure9 in https://arxiv.org/abs/2403.14112
PYTHONPATH=. python tools/summarize_mem_rea.py ${REASON_SUMMARY_CSV} ${MEMORY_SUMMARY_CSV}
# generate Table7, Table12, Table13 and Figure11 in https://arxiv.org/abs/2403.14112
PYTHONPATH=. python tools/analyze_mem_indep_rea.py data/CHARM ${REASON_RESULTS_DIR} ${MEMORY_SUMMARY_DIR} ${MEMORY_SUMMARY_CSV}
🖊️ Citation
@misc{sun2024benchmarking,
title={Benchmarking Chinese Commonsense Reasoning of LLMs: From Chinese-Specifics to Reasoning-Memorization Correlations},
author={Jiaxing Sun and Weiquan Huang and Jiang Wu and Chenya Gu and Wei Li and Songyang Zhang and Hang Yan and Conghui He},
year={2024},
eprint={2403.14112},
archivePrefix={arXiv},
primaryClass={cs.CL}
}