| --- |
| license: cc-by-4.0 |
| task_categories: |
| - text-generation |
| language: |
| - en |
| - zh |
| - es |
| - fr |
| - de |
| - ru |
| - ja |
| - th |
| - sw |
| - te |
| - bn |
| - ar |
| - ko |
| - vi |
| - cs |
| - hu |
| - sr |
| multilinguality: |
| - multilingual |
| size_categories: |
| - 1K<n<10K |
| configs: |
| - config_name: en |
| data_files: arenahard_en.jsonl |
| - config_name: zh |
| data_files: arenahard_zh.jsonl |
| - config_name: es |
| data_files: arenahard_es.jsonl |
| - config_name: fr |
| data_files: arenahard_fr.jsonl |
| - config_name: de |
| data_files: arenahard_de.jsonl |
| - config_name: ru |
| data_files: arenahard_ru.jsonl |
| - config_name: ja |
| data_files: arenahard_ja.jsonl |
| - config_name: th |
| data_files: arenahard_th.jsonl |
| - config_name: bn |
| data_files: arenahard_bn.jsonl |
| - config_name: sw |
| data_files: arenahard_sw.jsonl |
| - config_name: te |
| data_files: arenahard_te.jsonl |
| - config_name: ar |
| data_files: arenahard_ar.jsonl |
| - config_name: ko |
| data_files: arenahard_ko.jsonl |
| - config_name: vi |
| data_files: arenahard_vi.jsonl |
| - config_name: cs |
| data_files: arenahard_cs.jsonl |
| - config_name: hu |
| data_files: arenahard_hu.jsonl |
| - config_name: sr |
| data_files: arenahard_sr.jsonl |
| tags: |
| - multilingual |
| - instruction-following |
| --- |
| ## Dataset Sources |
|
|
| - **Paper**: BenchMAX: A Comprehensive Multilingual Evaluation Suite for Large Language Models |
| - **Link**: https://huggingface.co/papers/2502.07346 |
| - **Repository**: https://github.com/CONE-MT/BenchMAX |
|
|
| ## Dataset Description |
| BenchMAX_Model-based is a dataset of [BenchMAX](https://arxiv.org/pdf/2502.07346), sourcing from [m-ArenaHard](https://huggingface.co/datasets/CohereForAI/m-ArenaHard), which evaluates the instruction following capability via model-based judgment. |
| |
| We extend the original dataset to include languages that are not supported by m-ArenaHard through Google Translate. |
| Then manual post-editing is applied for all non-English languages. |
| |
| ## Usage |
| |
| ```bash |
| git clone https://github.com/CONE-MT/BenchMAX.git |
| cd BenchMAX |
| pip install -r requirements.txt |
| |
| cd tasks/arenahard |
| bash prepare.sh |
| ``` |
| |
| Then modify the model configs in `arena-hard-auto/config`. |
| Please add your model config to `api_config.yaml` and add your model name to the model list in other configs like `gen_answer_config_*.yaml`. |
| If you want to change the judge model, you can modify `judge_config_*.yaml`. |
| |
| Finally, deploy your model and run the evaluation, where your model first generates responses to prompts and DeepSeek-V3 judge them against GPT-4o responses, as we do in the paper. |
| |
| ```bash |
| # serve your model by vllm |
| vllm serve meta-llama/Llama-3.1-8B-Instruct |
| |
| # generate responses |
| cd arena-hard-auto |
| languages=(en ar bn cs de es fr hu ja ko ru sr sw te th vi zh) |
| for lang in "${languages[@]}"; do |
| python gen_answer.py --setting-file config/gen_answer_config_${lang}.yaml |
| done |
| |
| # run LLM-as-a-judge |
| export OPENAI_API_KEY=... |
| for lang in "${languages[@]}"; do |
| python gen_judgment.py --setting-file config/judge_config_${lang}.yaml |
| done |
| ``` |
| |
| ## Supported Languages |
| Arabic, Bengali, Chinese, Czech, English, French, German, Hungarian, Japanese, Korean, Serbian, Spanish, Swahili, Telugu, Thai, Russian, Vietnamese |
| |
| ## Citation |
| If you find our dataset helpful, please cite this paper: |
| |
| ``` |
| @article{huang2025benchmax, |
| title={BenchMAX: A Comprehensive Multilingual Evaluation Suite for Large Language Models}, |
| author={Huang, Xu and Zhu, Wenhao and Hu, Hanxu and He, Conghui and Li, Lei and Huang, Shujian and Yuan, Fei}, |
| journal={arXiv preprint arXiv:2502.07346}, |
| year={2025} |
| } |
| ``` |
