physcorp-pre-audit
收藏资源简介:
PhysCorp Pre-Audit Raw Pool 是一个包含14,294条记录的物理推理多模态语料库,作为论文《Physics-R1: An Audited Olympiad Corpus and Recipe for Visual Physics Reasoning》的伴随资源发布。它聚合了来自九个不同来源的原始数据,形成在应用联合污染审核之前的完整语料库,核心目的是允许用户基于此原始池以不同阈值或针对新外部基准重新运行审核流程。数据集涵盖物理学科,特别侧重于奥林匹克竞赛级别的推理问题,任务类别包括图像-文本到文本和问答。数据形式包含文本和多模态样本,其中1,415条记录为多模态行,在`messages`字段中包含`<image>`标记,多模态图像主要来自MMMU Physics公共基准和PhysReason数据集。数据结构遵循固定schema,包含`source`(数据来源)、`messages`(JSON序列化的消息/问题)、`solution`(解决方案)和`metadata`(JSON序列化的元数据)字段。数据来源多样,包括UGPhysics、OpenStax College + University Physics等,每个来源都有明确的记录数、许可证(如CC BY-NC-SA 4.0、CC BY 4.0、公共领域等)和质量层级(Tier-1或Tier-2)标注。该预审核池是后续审核流程的起点:通过两阶段联合污染审核(第一阶段5-gram Jaccard >= 0.4,第二阶段mxbai-embed-large-v1余弦相似度 >= 0.85)针对六个标准评估集进行筛选,最终产生6,432条的审核后语料库(physcorp-a)和2,268条的封闭形式强化学习训练池(physr1corp)。数据集整体采用CC BY-NC 4.0许可证,但用户可根据`source`列中的信息按各来源原始许可证进行下游过滤和使用。
PhysCorp Pre-Audit Raw Pool is a multimodal corpus for physics reasoning containing 14,294 records. It is released as a companion resource to the paper Physics-R1: An Audited Olympiad Corpus and Recipe for Visual Physics Reasoning, aiming to aggregate raw data from nine different sources into a complete corpus before applying joint contamination auditing. Its core purpose is to allow users to re-run the auditing process on this raw pool with different thresholds or for new external benchmarks. The dataset covers the physics discipline, with a particular focus on Olympiad-level reasoning problems, and task categories include image-text-to-text and question answering. The data format includes text and multimodal samples, with 1,415 records being multimodal rows containing `<image>` tags in the `messages` field. These multimodal images primarily come from the MMMU Physics public benchmark and the PhysReason dataset. The dataset structure follows a fixed schema with four fields: `source` (data source), `messages` (JSON-serialized messages/questions), `solution` (solution), and `metadata` (JSON-serialized metadata). Data sources are diverse, including UGPhysics, OpenStax College + University Physics, Physics Stack Exchange, Kevin Zhous Olympiad handouts, Estonian Physics Olympiad (EFO), MMMU + o1-CoT seed data, PhysReason, and data scraped from competitions such as IPhO, NBPhO, EuPhO, APhO, USAPhO, and INPhO, each annotated with clear record counts, licenses (e.g., CC BY-NC-SA 4.0, CC BY 4.0, public domain), and quality tiers (Tier-1 or Tier-2). This pre-audit pool serves as the starting point for subsequent auditing: it undergoes a two-stage joint contamination audit (stage 1: 5-gram Jaccard >= 0.4, stage 2: mxbai-embed-large-v1 cosine similarity >= 0.85) against six standard evaluation sets, ultimately producing an audited corpus of 6,432 records (physcorp-a) and a closed-form reinforcement learning training pool of 2,268 records (physr1corp). The overall dataset is licensed under CC BY-NC 4.0, but users can filter and use it downstream based on the original licenses of each source as indicated in the `source` column.
数据集概述
PhysCorp Pre-Audit Raw Pool 是一个未经污染审计的物理推理原始语料库,包含 14,294条记录,是 Physics-R1 项目的一部分,旨在为用户提供可自行调整阈值或针对新基准重新运行污染审计的灵活性。
基本信息
- 许可证: CC BY-NC 4.0
- 语言: 英语、中文
- 任务类型: 图像-文本到文本、问答
- 标签: 物理、推理、多模态、奥林匹克竞赛、预审计
- 数据规模: 10K < n < 100K
- 配置: 默认配置,包含一个训练集文件(
physcorp_pre_audit.parquet)
数据来源与构成
数据聚合自九个来源家族,记录数及许可证信息如下:
| Source | Records | License | Tier |
|---|---|---|---|
| UGPhysics | 5,520 | CC BY-NC-SA 4.0 | Tier-2 |
| OpenStax College + University Physics | 2,381 | CC BY 4.0 | Tier-1 |
| Physics Stack Exchange | 2,291 | CC BY-SA 4.0 | Tier-1 |
| Kevin Zhous olympiad handouts | 692 | CC BY-NC 4.0 | Tier-2 |
| Estonian Physics Olympiad (EFO) | 418 | Public-domain (competition policy) | Tier-2 |
| MMMU + o1-CoT seed | 1,293 | MIT (MMMU); generated CoT | Tier-1 |
| PhysReason | 1,200 | CC BY 4.0 | Tier-1 |
| IPhO + NBPhO + EuPhO scrape | 258 | Public-domain | Tier-1 |
| APhO + USAPhO + INPhO scrape | 241 | Public-domain | Tier-1 |
| Total | 14,294 |
数据结构与加载
- 模式: 与
physcorp-a相同,包含字段:source,messages,solution,metadata。其中messages和metadata以 JSON 格式序列化存储于 parquet 文件中。 - 多模态记录: 包含 1,415条 多模态行(
messages中含有<image>标记)。其中:- 441条 来自 MMMU Physics(
source = rl_sft_*),可通过提供的loader.py脚本解析图像路径。 - 974条 来自 PhysReason(
source = physreason_*),需用户自行运行 代码仓库 中的prepare_additional_training_data.py来获取图像,原因在于上游 PhysReason 数据集的分发条款限制。
- 441条 来自 MMMU Physics(
- 图像路径映射: 存储在
image_mapping.json文件中(source_id → filename)。extra_images/目录下的438张PNG图像来自 MMMU 公共基准。
使用示例
-
基础加载: python from datasets import load_dataset ds = load_dataset("shanyangmie/physcorp-pre-audit", split="train")
-
多模态记录加载(使用脚本): python from loader import load_physcorp_pre_audit df = load_physcorp_pre_audit()
-
命令行解析: bash python loader.py --out resolved.parquet
污染审计流程
该数据集是污染审计前的原始池,经过两阶段联合审计(Stage-1: 5-gram Jaccard >= 0.4;Stage-2: mxbai-embed-large-v1 cosine >= 0.85)后,可得到更小的子集:
- 14,294条 预审计池(当前数据集)
- → 6,432条 审计后子集:
shanyangmie/physcorp-a - → 2,268条 闭式RL池:
shanyangmie/physr1corp(实际用于训练)
相关数据集
shanyangmie/physcorp-a— 7,236条记录的审计后语料库shanyangmie/physr1corp— 2,268条记录的闭式RL池shanyangmie/physolym-a— 500条记录的审计后奥林匹克评测集
许可证与引用
- 许可证: CC BY-NC 4.0,每条记录在
source列中提供来源归属,便于按许可证层级过滤。 - 引用: 如需引用,请使用提供的 BibTeX 格式。





