Skip to content

Latest commit

 

History

History
119 lines (85 loc) · 3.96 KB

README.md

File metadata and controls

119 lines (85 loc) · 3.96 KB

Open-Ended LLM Judgement

  • We currently support three benchmarks
    • MT-Bench
    • Vicuna benchmark
    • Advbench

You could firstly run utils/merge_lora.py to merge LORA to the base model.

MT-Bench

MT-Bench is an evaluation on two-turn conversations from FastChat.

Step 1. Generate model answers to MT-bench questions

python gen_model_answer_mt.py --base_model_path [MODEL_PATH] --template [TEMPLATE] --lora_path [LORA_PATH]

If you have merged LORA, you do not need to pass [LORA_PATH].

The answers will be saved to data/mtbench/model_answer/[MODEL-ID].jsonl, where [MODEL-ID] is extracted from the path name.

Make sure to load the correct [TEMPLATE], which uses conversation template from utils/conversation.py. Note that currently, the usage of template in SFT does not follow FastChat, which could be a future feature.

Step 2. Generate judgments

We use single-answer grading setting here, where GPT-4 directly give a score on a scale of 10.

export OPENAI_API_KEY=XXXXXX  # set the OpenAI API key
python gen_judge_mtbench.py --judge_model gpt-4-1106-preview --model_list [LIST-OF-MODEL-ID] --parallel [num-concurrent-api-call]   # GPT-4-Turbo

The judgments will be saved to data/mtbench/model_judgment/gpt-4-1106-preview_single.jsonl

Step 3. Show MT-bench scores

  • Show the scores for selected models 
    python show_results_mt.py --model_list [LIST-OF-MODEL-ID] --judge_model gpt-4-1106-preview
  • Show all scores 
    python show_results_mt.py

Vicuna and AdvBench

  • Vicuna benchmark is an evaluation on single-turn conversation from FastChat.
  • AdvBench is an attack evaluation benchmark that consists of 520 harmful questions, which is from llm-attacks.

Step 1. Generate model answers

If you do not merge LORA:

python gen_model_answer.py --base_model_path [MODEL_PATH] --template [TEMPLATE] --lora_path [LORA_PATH] --bench_name [BENCH_NAME]

If you have merged LORA:

python gen_model_answer.py --base_model_path [MODEL_PATH] --template [TEMPLATE] --bench_name [BENCH_NAME] --use_vllm

--use_vllm is not a must, but it will be extremely faster for inference.

The answers will be saved to data/[BENCH_NAME]/model_answer/[MODEL-ID].json, where [MODEL-ID] is extracted from the path name.

Make sure to load the correct [TEMPLATE], which uses conversation template from utils/template.py.

Step 2. Generate judgments

Generate judgment on Vicuna benchmark

export OPENAI_API_KEY=XXXXXX  # set the OpenAI API key
python gen_judge_vicuna.py --model_answer [MODEL-ID] --judger gpt-4

The judgments will be saved to data/vicuna/model_judgment/[JUDGER]_[MODEL-ID].json

Generate judgment on Advbench

python gen_judge_advbench.py --model_answer [MODEL-ID] --judger rule  # rule-based evaluation

The judgments will be saved to data/advbench/model_judgment/[JUDGER]_[MODEL-ID].json. It will directly output the final judgment result.

Step 3. Show Vicuna scores

python show_results_vicuna.py --eval_list [EVAL-LIST-ID]

Citation

For MT-Bench and Vicuna Benchmark:

@misc{zheng2023judging,
      title={Judging LLM-as-a-judge with MT-Bench and Chatbot Arena},
      author={Lianmin Zheng and Wei-Lin Chiang and Ying Sheng and Siyuan Zhuang and Zhanghao Wu and Yonghao Zhuang and Zi Lin and Zhuohan Li and Dacheng Li and Eric. P Xing and Hao Zhang and Joseph E. Gonzalez and Ion Stoica},
      year={2023},
      eprint={2306.05685},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

For Advbench:

@misc{zou2023universal,
      title={Universal and Transferable Adversarial Attacks on Aligned Language Models}, 
      author={Andy Zou and Zifan Wang and J. Zico Kolter and Matt Fredrikson},
      year={2023},
      eprint={2307.15043},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}