fineproofs-prm-context-v2-full-cot-xprob
收藏资源简介:
该数据集是FineProofs PRM Context v2系列的一个变体,名为Full Cot Xprob。它专为过程奖励模型(Process Reward Model)训练和定理证明任务设计,属于文本分类数据集。数据集基于已验证的FineProofs回滚集合构建,采用“完整思维链(full_cot)”上下文和“跨问题(cross_problem)”回滚策略,并使用了“中间截断推理等分共享(middle_truncated_reasoning_equal_share)”打包方式。每个样本包含部分前缀和完整响应目标,均使用标准化的规范评分信用(即钳制点数除以最大点数)进行标注。训练目标为密集的“奖励(reward)”值,而“正确(correct)”列是基于奖励≥0.5的布尔投影,仅供遗留参考。数据集包含训练集(53,457行,2,542个问题)和验证集(2,602行,128个问题),所有样本具有相同的行键、标签和奖励分布。该变体与其他八个上下文变体共享相同的问题划分和硬端点覆盖,未包含的问题在分割前已被移除。数据以Parquet格式存储,并附有来源证明、验证信息和文件指纹等元数据。
This dataset is a variant of the FineProofs PRM Context v2 series, named Full Cot Xprob. It is designed for Process Reward Model training and theorem proving tasks, belonging to a text classification dataset. The dataset is built on the validated FineProofs rollback collection, employing a full_cot context and cross_problem rollback strategy, along with a middle_truncated_reasoning_equal_share packing method. Each sample contains a partial prefix and a complete response target, both annotated with a standardized normalized scoring credit (i.e., clamped points divided by maximum points). The training target is a dense reward value, while the correct column is a boolean projection based on reward ≥ 0.5, provided for legacy reference. The dataset includes a training set (53,457 rows, 2,542 problems) and a validation set (2,602 rows, 128 problems), with all samples sharing the same row keys, labels, and reward distribution. This variant shares the same problem split and hard endpoint coverage with eight other context variants; problems not included were removed before splitting. The data is stored in Parquet format, with metadata such as source provenance, verification information, and file fingerprints.
数据集概述
FineProofs PRM Context v2 - Full Cot Xprob 是一个用于过程奖励模型(Process Reward Model, PRM)训练与评估的文本分类数据集,专注于定理证明(theorem-proving)任务,基于 FineProofs 验证过的 rollout 集合构建。
核心特性
- 上下文模式(Context Mode):采用
full_cot,即使用完整的思维链(Chain-of-Thought)作为上下文。 - 上下文范围(Context Scope):
cross_problem,即跨问题(cross_problem)的 rollout 数据;打包策略为middle_truncated_reasoning_equal_share。 - 标签说明:
correct列仅为旧版布尔投影(奖励 ≥ 0.5 时为 True);正式训练使用稠密奖励列reward,该奖励基于规范化后的标准评分(规范化积分除以最大积分)计算。 - 验证信息:校验 SHA-256 为
c5a3b92bd6196d4a33cded2c1a01870d1963d323351d573b369b7f97f3b7f648。
数据集构成
该数据集是九个行匹配的上下文变体之一。所有变体共享相同的行键、标签、奖励、训练/验证问题划分以及硬端点覆盖;已排除留出(held-out)问题后再进行划分和上下文选择。
| 划分 | 行数 | 问题数 |
|---|---|---|
| train | 53,457 | 2,542 |
| validation | 2,602 | 128 |
元数据与合约信息
- 运行 ID:
fineproofs_all_qwen35_9b_direct2phase_m32_20260730 - 采集模型:
Qwen/Qwen3.5-9B - 采集模型修订版本:
c202236235762e1c871ad0ccb60c8ee5ba337b9a - 上下文变体名称:
full_cot_xprob - 上下文模式:
full_cot - 上下文范围:
cross_problem - 上下文中的正确性标签:不含(
no)
此外,发布文件中包含 dataset_provenance.json、verification.json 和 _SUCCESS.json,分别用于记录实时输入来源、跨变体检查以及已发布文件的精确指纹。




