---
license: cc0-1.0
pretty_name: SIRIS Academic Affiliation Benchmark
language:
- multilingual
configs:
- config_name: datacite
  data_files:
  - split: train
    path: data/datacite/train-00000-of-00001.parquet
- config_name: arxiv
  data_files:
  - split: train
    path: data/arxiv/train-00000-of-00001.parquet
---

# SIRIS academic affiliation benchmark

This dataset contains two samples of 3,000 distinct academic
affiliation strings. Each configuration has one `affiliation` string column.
It contains no ROR labels or source metadata columns.

## DataCite configuration

The population is 1,681,662 exact rows from the `train` split of
`cometadata/datacite-unique-affiliations` at revision
`e7606edb694d3dcd4a7ea86a146818bd5a339738`. Sampling is without replacement
and weighted by `occurrence_count`. For a seeded SHA-256 uniform value `U`, the
priority is `-log(U) / occurrence_count`; the 3,000 lowest priorities
are selected.

## arXiv configuration

The population is 2,399,952 exact distinct strings extracted from
`prediction[].affiliations[].affiliation` in
`cometadata/arxiv-author-affiliations-matched-ror-ids`. The source main revision
is `a56eb960359fb8dc32ee33107e9e501ddd7e9c3d`; processing uses immutable
Parquet conversion revision `97a71fa6d850d11af13a0f2c98d0f1f206b44955`.
The 3,000 lowest seeded SHA-256 values are selected uniformly without
replacement.

## Exact-text and reproducibility policy

Both samples use seed `42` and algorithm identifier
`siris-affiliation-sample-v1`. Strings retain their original case and all
whitespace. Null arXiv values are excluded; empty and whitespace-only strings
remain eligible. No string is trimmed, normalized, transliterated, or rewritten.

## License

Both sources and this derived dataset are released under CC0 1.0.