ScienceSoft/piibench

Tasks

• token-classification

Citation text from the card

piibench is free to use, and crediting ScienceSoft is a condition of that use, not a courtesy.
Three layers, because three different things are being licensed.

**ScienceSoft's own work: CC BY 4.0 — attribution required.** The curation of each dataset's
gold into the scored and regex-owned sides, the windowing, the per-set balancing, the carrier
serialisation, the metric and scoring code, the published results in `results/`, and the
leaderboard. CC BY 4.0 makes attribution a licence condition: use it, adapt it, sell a product
informed by it — and credit the source. Attribute as:

> piibench, ScienceSoft — https://huggingface.co/datasets/ScienceSoft/piibench

**The framework code: Apache-2.0, with a NOTICE.** The code under `src/`, `scripts/` and
`tests/` stays Apache-2.0 so it can be vendored without friction. `NOTICE` is the file Section
4(d) of that licence makes binding: **anyone redistributing this code, or a derivative of it,
must include it.** That is the attribution hook for the code path.

**The upstream rows: their own authors' terms, which we do not extend.** The rows in
`data/rows.jsonl` are derived from seven third-party datasets, and each set's rows carry the
licence of the source they came from: MIT (TAB, FactRuEval), CC BY 4.0 (MEDDOCAN, IAHLT,
Nemotron), CC BY-SA 4.0 (GermEval — ShareAlike), Apache-2.0 (Gretel).

**We cannot require ScienceSoft attribution for those, and do not claim to.** They are other
people's work, and CC BY-SA in particular forbids imposi

All datasets marked Permissive or share-alike | JSON