PhysCorp-A: An Audited Physics Training Corpus

6,432 records form an audited physics training corpus released alongside the paper 'Physics-R1: An Audited Olympiad Corpus and Recipe for Visual Physics Reasoning'. The dataset is a subset of a 14,294-record pre-audit pool that survived a two-stage contamination audit against six canonical evaluation splits. It was created by shanyangmie and last updated on Hugging Face in May 2026.

Use Cases

• Training visual physics reasoning models based on the described audited corpus.
• Evaluating model performance on physics problems using the audit-cleaned training data.
• Studying the effects of data contamination and audit processes on model training.
• Benchmarking AI systems against Olympiad-level physics reasoning tasks.

Strengths

• 6,432 records provide a defined scale for training.
• Dataset is the result of a two-stage contamination audit against six evaluation splits.
• Derived from a larger pre-audit pool of 14,294 records.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is known, but the specific features and data format are unknown.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

Hugging Face dataset by shanyangmie, associated with the 'Physics-R1' research paper.

Collection Method

Subset of a 14,294-record pre-audit pool filtered through a joint two-stage contamination audit.

Freshness

Last updated 2026-05-19 05:30:17; freshness should be verified.