fineproofs-prm-context-v2-solution-label
收藏资源简介:
该数据集是FineProofs PRM Context v2的Solution Label分支,专为过程奖励模型(Process Reward Model, PRM)训练设计,属于文本分类任务。数据来源于FineProofs的已验证回滚集合,采用same_problem策略和solution_label上下文,打包策略为whole_solutions_no_truncation。每个样本包含部分前缀或完整响应,其奖励目标基于规范化评分(夹紧点数除以最大点数),correct列为奖励≥0.5的布尔投影。训练集包含53,457行,对应2,542个问题;验证集包含2,602行,对应128个问题。该数据集适用于定理证明中的过程奖励建模,支持密集奖励训练。
This dataset is the Solution Label branch of FineProofs PRM Context v2, specifically designed for training Process Reward Models (PRM) as a text classification task. It is sourced from the validated rollback collection of FineProofs, adopting the same_problem strategy and solution_label context, with the packing strategy set to whole_solutions_no_truncation. Each sample contains partial prefixes or full responses, and its reward target is based on normalized scores (clamped points divided by the maximum points). The "correct" column is a boolean projection where the reward value is ≥ 0.5. The training set includes 53,457 rows corresponding to 2,542 problems, while the validation set contains 2,602 rows corresponding to 128 problems. This dataset is applicable to process reward modeling in theorem proving and supports dense reward training.
数据集概述:FineProofs PRM Context v2 - Solution Label
核心信息
- 名称:FineProofs PRM Context v2 (Solution Label)
- 任务类别:文本分类(text-classification)
- 标签:过程奖励模型(process-reward-model)、定理证明(theorem-proving)、FineProofs
- 数据格式:Parquet 文件,包含训练集(train.parquet)与验证集(validation.parquet)
数据来源与构建方式
- 该数据集是 FineProofs 验证过的 rollout 集合中九个上下文变体之一,采用 solution_label 上下文,来自 same_problem rollouts,打包策略为 whole_solutions_no_truncation。
- 运行 ID:
fineproofs_all_qwen35_9b_direct2phase_m32_20260730 - 采集模型:
Qwen/Qwen3.5-9B(修订版c202236235762e1c871ad0ccb60c8ee5ba337b9a) - 上下文模式与范围:solution_label / same_problem
- 上下文中的正确性标签:有(yes)
- 验证哈希(SHA-256):
c5a3b92bd6196d4a33cded2c1a01870d1963d323351d573b369b7f97f3b7f648
标签与目标
- 部分前缀和完整响应目标均使用规范化的归一化积分(clamped points / max points)。
correct列仅为旧式布尔投影(reward >= 0.5),训练时使用稠密reward目标。
数据划分
| 划分 | 行数 | 问题数 |
|---|---|---|
| train | 53,457 | 2,542 |
| validation | 2,602 | 128 |
- 九个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。
- 在划分和上下文选择之前,已移除保留问题(held-out problems)。
- 本数据集提供
dataset_provenance.json、verification.json和_SUCCESS.json,分别包含实时输入来源、跨变体验证和发布文件精确指纹。



