fusion-pairwise-evals-finetuned

# Automatic pairwise preference evaluations for: Making, not taking, the Best-of-N ## Content This data contains pairwise automatic win-rate evaluations for the [m-ArenaHard-v2.0](https://huggingface.co/datasets/CohereLabs/m-ArenaHard-v2.0) benchmark and it compares 2 models against gemini-2.5-flash: 1. `Fusion`: is the 111B model finetuned on [synthetic data](https://huggingface.co/datasets/CohereLabs/fusion-synth-data-ufb) generated with Fusion from 5 teachers 2. `BoN`: is the 111B model finetuned on [synthetic data](https://huggingface.co/datasets/CohereLabs/fusion-synth-data-ufb) generated with BoN from 5 teachers Each model’s outputs are compared in pairs with the respective Gemini output, and judged by GPT-4o. For an analysis and context of these evaluations, check out the [paper](https://arxiv.org/abs/2510.00931). ## Format The data is organized in Jsonlines format where each line contains all the information for a single prompt evaluation. Below we explain the format for a sample line, annotations in "<>": ``` { question_id: <the unique ID for the question in m-ArenaHard-v2.0 for the example (same across languages)> prompt: <text of this prompt> language_code: <Language code of this prompt> BoN: <The judge outcome for BoN finetuned model generation for this prompt. one of three values {win, loss, tie}> Fusion: <The judge outcome for Fusion finetuned model generation for this prompt. one of three values {win, loss, tie}> bon_completion: <The BoN finetuned model generation for this prompt> fusion_completion: <The Fusion finetuned model generation for this prompt> gemini_2_5_flash_completion: <Gemini-2.5-flash generation for this prompt> judge_model_name: <Name of the judge model used for the pairwise comparison. This is always gpt-4o-2024-05-13> } ``` ## Use **This data may not be used for model training!** You may use this data to conduct analyses of model differences, evaluate other judges against GPT4o, or similar inference-only experiments. Make sure to also respect the individual licenses for using outputs from Gemini, GPT4o, and CommandA models. ## Citation If you use this data for your research, please cite our work accordingly: ``` @misc{khairi2025makingtakingbestn, title={Making, not Taking, the Best of N}, author={Ammar Khairi and Daniel D'souza and Marzieh Fadaee and Julia Kreutzer}, year={2025}, eprint={2510.00931}, archivePrefix={arXiv}, primaryClass={cs.CL}, url={https://arxiv.org/abs/2510.00931}, } ```