PhysCISC: Code and dataset for process-credibility weighted self-consistency voting in physics reasoning

PhysCISC (v1.0.0): A Physics Multiple-Choice Reasoning Benchmark and Reproducible Codebase for Process-Credibility-Weighted Self-Consistency (PhysCISC) This Zenodo record contains both the PhysCISC dataset and the full codebase used to (i) construct the benchmark and (ii) run the SC/CISC/PhysCISC experiments reported in our paper. The package is released to support reproducibility and to enable future research on test-time aggregation / self-consistency and process-quality signals for LLM physics reasoning. What’s included Dataset (PhysCISC): 4,078 physics multiple-choice questions spanning 6 domains and 5 reasoning types, with metadata for domain/type tags. Reproducible experiments: scripts/configs for SC, CISC (P(True) weighting), and PhysCISC (process-credibility weighting), including evaluation utilities. Dataset construction pipeline: code used for parsing/structuring, filtering, converting to MCQ format, and tagging. Documentation: dataset schema, statistics, release notes, checksums, and a quick “how to reproduce” guide. Dataset format and fields The public release provides the dataset in machine-readable formats (JSON and JSONL) together with: a stable question ID for each item, domain/type annotations, answer key and minimal metadata needed to reproduce the evaluation protocol, a JSON schema and basic statistics files. Copyright/redistribution note: the released dataset does not include scanned pages or restricted verbatim reproductions of copyrighted books. It contains the released benchmark text + metadata intended for research use. Reproducibility (quick start) Download and unzip this record. Follow README.md for environment setup. Use the provided scripts to reproduce: SC majority voting, CISC with confidence signal P(True), PhysCISC with process-credibility scoring and softmax-temperature weighting. License Data: CC BY 4.0 (attribution required). Code: Apache 2.0.(See LICENSE-DATA and LICENSE-CODE in the archive.) Zenodo requires selecting a license at upload time; CC BY 4.0 is commonly used for research data.Zenodo Citation Please cite this Zenodo record if you use the dataset or code. Zenodo uses version DOIs and a concept DOI that groups all versions.Zenodo+1