mmlu-pro-olmo-3-7b-instruct-temp0.9-samples99-logprobs
收藏资源简介:
该数据集包含针对MMLU-Pro测试集的99个自一致性生成结果,每个问题均附带完成时的令牌级对数概率。数据由`allenai/OLMo-3-7B-Instruct`模型在温度参数`0.9`下生成,适用于事后分析、自一致性曲线、自适应停止及相关聚合方法。数据集来源包括基准测试`TIGER-Lab/MMLU-Pro`和模型`allenai/OLMo-3-7B-Instruct`,解码参数为`num_samples=99`和`top_p=0.8`。数据集文件为Parquet格式,包含问题ID、问题内容、选项、可能答案、所有完成结果、最终答案及所有对数概率。数据集的准确率、F1分数和精确匹配率均为0.596908244680851。
This dataset contains 99 self-consistency generation results for the MMLU-Pro test set, with token-level log probabilities of the completion attached to each question. Generated by the `allenai/OLMo-3-7B-Instruct` model with a temperature parameter of 0.9, this dataset is suitable for post-hoc analysis, self-consistency curves, adaptive stopping, and related aggregation methods. The dataset sources include the benchmark `TIGER-Lab/MMLU-Pro` and the model `allenai/OLMo-3-7B-Instruct`, with decoding parameters set to `num_samples=99` and `top_p=0.8`. The dataset files are in Parquet format, containing question ID, question content, options, candidate answers, all completion results, final answers, and all log probabilities. The accuracy, F1 score, and exact match rate of this dataset are all 0.596908244680851.




