hotpotqa-dev-olmo-3-7b-instruct-temp0.9-samples99-logprobs
收藏资源简介:
该数据集包含针对HotpotQA验证集的每个问题生成的99个自一致性完成样本,使用`allenai/OLMo-3-7B-Instruct`模型在温度为`0.9`的条件下生成,并附有每个完成样本的令牌级对数概率。数据集旨在用于事后分析、自一致性曲线、自适应停止及相关聚合方法的研究。数据来源于`hotpotqa/hotpot_qa`基准测试的`distractor`和`validation`部分,采用温度采样解码方式,参数为`num_samples=99`和`top_p=0.8`,并通过事后HotpotQA答案匹配进行评分。数据集包含多个文件,其中parquet文件包含`question`、`question_id`、`possible_answers`、`all_completions`、`final_answer`和`all_logprobs`等字段。此外,README还提供了不同聚合变体的性能比较,显示归一化多数投票方法表现最佳。
This dataset contains 99 self-consistency completion samples generated for every question in the HotpotQA validation set. These samples were produced using the `allenai/OLMo-3-7B-Instruct` model with a temperature set to 0.9, and each sample is accompanied by its token-level log probabilities. The dataset is designed for research on post-hoc analysis, self-consistency curves, adaptive stopping, and related aggregation methodologies. It is sourced from the `distractor` and `validation` splits of the `hotpotqa/hotpot_qa` benchmark, and was generated via temperature-sampling decoding with parameters `num_samples=99` and `top_p=0.8`. All samples are scored through post-hoc HotpotQA answer matching. The dataset includes multiple files, with the Parquet files containing fields including `question`, `question_id`, `possible_answers`, `all_completions`, `final_answer`, and `all_logprobs`. Additionally, the accompanying README provides performance comparisons across different aggregation variants, showing that the normalized majority voting method delivers the best performance.




