| # Attribution And Notice Of Adaptation | |
| ## FindStat | |
| FindStatBench is derived from **FindStat - the combinatorial statistics database**, maintained by Martin Rubey, Christian Stump and the FindStat contributors: https://www.findstat.org/. | |
| Historical project/license reference: Chris Berg, Viviane Pons, Travis Scrimshaw, Jessica Striker and Christian Stump (2014), *FindStat - the combinatorial statistics database*, arXiv:1401.3690v1, https://arxiv.org/abs/1401.3690. Page 2 and footnote 4 identify CC BY-SA 3.0 for database/wiki contributions. See `DATA_RIGHTS.md` for the historical scope qualification. | |
| Source descriptions, mathematical encodings, examples and labels retain their upstream origin. Decoded task records preserve original source IDs, contributor/update metadata, references and source-relative provenance paths. Individual entries can be referenced using their original `St`/`Mp` identifiers at FindStat. Follow the database citation instructions at https://www.findstat.org/Citations/. | |
| ## FindStatBench Adaptations | |
| FindStatBench benchmark curation/packaging: **Soham Dan**. Source content was organized into statistic- and map-synthesis tasks, public example and evaluation-instance records, and versioned prompt exports. The historical closed-book prompt transformation removes direct source-identification headers but does not eliminate every cross-reference or source link. | |
| The Hugging Face packaging converts the frozen JSONL records into standard Parquet shards and encodes heterogeneous nested fields as JSON strings. The supplied decoder reconstructs the complete logical records without changing values, types or order. No mathematical task, gold label or historical served prompt was added, removed or corrected during this conversion. The single `test` split contains the complete corpus rather than imposing a training partition. | |
| These adaptations are identified to distinguish FindStatBench from the upstream database. **No FindStat contributor endorsement or sponsorship is implied.** Data adaptations are offered under CC BY-SA 3.0 to the extent of the applicable upstream grant and the maintainer's rights; full terms are in `LICENSE`. Software is separately MIT-licensed in `LICENSE-CODE.txt`. | |