---
configs:
- config_name: default
  data_files:
  - split: train
    path: train.jsonl
  - split: gold
    path: gold.jsonl
license: apache-2.0
---

# data-use-evidence-classifier-data

Training corpus for the evidence-tier classifier (v3): 62,054 data-use mention
spans (marked `>>> mention <<<` in context) with luna-judged tiers
(tier1_evidential / tier2_declaration / tier3_nonmention / junk), distilled from
`rafmacalaba/data-use-mentions-v2` (val + holdout dual-pass + train-top single-pass).

- `train` — 62,054 spans, decontaminated against the gold controls
- `gold` — 573 human-adjudicated controls (573 gold tier labels), held out for benchmarking

Source: `rafmacalaba/data-use-mentions-v2` (gliner2_v2 `output_meta` overlay).