CoolFace
Datasetpublic

rafmacalaba/data-use-mentions-tiered

Data-use mentions — tiered copy (T1∪T2-only supervision) Derived from rafmacalaba/data-use-mentions (originals untouched). Same windows, same text; spans judged tier3_nonmention or junk (Luna verdicts; unjudged train spans via v3 tier classifier at p_t3+p_junk >= 0.9) are UNTAGGED — text stays, so they act as hard negatives for span-extraction training. Total untagged: 28807 spans. Labels: single DATA_MENTION class (kept spans = T1 evidential ∪ T2 declaration). Specificity… See the full description on the dataset page: https://huggingface.co/datasets/rafmacalaba/data-use-mentions-tiered.

sourceHugging Facecc-by-4.0updated 20d agoView on Hugging Face
0likes276downloads
Dataset Card

Data-use mentions — tiered copy (T1∪T2-only supervision)

Derived from rafmacalaba/data-use-mentions (originals untouched). Same windows, same text; spans judged tier3_nonmention or junk (Luna verdicts; unjudged train spans via v3 tier classifier at pt3+pjunk >= 0.9) are UNTAGGED — text stays, so they act as hard negatives for span-extraction training. Total untagged: 28807 spans.

Labels: single DATA_MENTION class (kept spans = T1 evidential ∪ T2 declaration). Specificity (named/descriptive/vague) is handled downstream by the multitask SFT model, not the extractor.

{"holdout": {"kept": 17416, "negative_only_newly_emptied": 3377, "negative_only_source": 3377, "negative_only_used": 3586, "negative_ratio": 0.3, "positive_rows": 11955, "rows_out": 15541, "source_rows": 18709, "source_spans": 22052, "unmatched_kept": 0, "untag_by_source": {"luna": 4636}, "untagged": 4636}, "train": {"kept": 84039, "negative_only_newly_emptied": 13770, "negative_only_source": 15324, "negative_only_used": 16772, "negative_ratio": 0.3, "positive_rows": 55909, "rows_out": 72681, "source_rows": 85003, "source_spans": 103230, "unmatched_kept": 0, "untag_by_source": {"classifier": 10723, "luna": 8468}, "untagged": 19191}, "val": {"kept": 18322, "negative_only_newly_emptied": 3519, "negative_only_source": 3308, "negative_only_used": 3697, "negative_ratio": 0.3, "positive_rows": 12325, "rows_out": 16022, "source_rows": 19152, "source_spans": 23302, "unmatched_kept": 0, "untag_by_source": {"luna": 4980}, "untagged": 4980}}

Usage:

python
from datasets import load_dataset
ds = load_dataset("rafmacalaba/data-use-mentions-tiered", "gliner_tiered")

Synthetic rows (is_synthetic: true)

Train-split rows generated locally with Gemma-4-E4B (tiered-anchored, invented mentions; full pipeline, gates, and quality stats: `docs/synthetic-generation.md` in the ai4data-playground repo).

  • —is_synthetic — false for the original corpus (origin prwp/fcv), true for generator output. Val/holdout are all-real.
  • —origin_model — the generator that produced a synthetic row (empty for real rows).
  • —Synthetic rows carry origin: "synth"; labels for usage-impact are copied from verified seed roles (never model-generated); provenance attributes are verbatim-gated; bibtex is assembled deterministically.
  • —Filter for real-only: ds.filter(lambda r: not r["is_synthetic"]).