---
task_categories:
  - token-classification
tags:
  - gliner
  - gliner2
  - information-extraction
  - probe-review-sample
license: cc-by-4.0
configs:
  - config_name: gliner
    data_files:
      - split: sample
        path: gliner_sample.jsonl
  - config_name: bio
    data_files:
      - split: sample
        path: bio_sample.jsonl
  - config_name: gliner2
    data_files:
      - split: sample
        path: gliner2_sample.jsonl
      - split: review
        path: gliner2_review.jsonl
---

# Probe-review sample of `rafmacalaba/datause-extracted-sample`

Stratified review slice: whole rows sampled per origin until each specificity reaches ~150 spans (seed 0; small origins contribute all they have), plus entity-less calibration rows round-robined across `neg_class` shapes. Rows keep their original `split` values. Score with training/score_extract_probe.py `--splits sample`.

| origin | rows | negatives | named / descriptive / vague spans |
|---|---|---|---|
| `fcv_pads_east_africa` | 389 | 60 | 171 / 187 / 150 |
| `general_prwp` | 317 | 60 | 192 / 187 / 150 |
| `jad_paddy_docs` | 140 | 7 | 187 / 63 / 46 |
| `jdc_operational` | 310 | 60 | 175 / 157 / 76 |
| `refugee_pads` | 376 | 60 | 169 / 182 / 150 |
| `reliefweb` | 339 | 60 | 166 / 170 / 150 |

## Probe scores

Every span carries `probe_score` (sigmoid head output ~= P(genuine data-use mention)) from `rafmacalaba/gliner_datause_probe` (frozen `rafmacalaba/gliner_datause` encoder, context_radius=64, threshold=None). Extractor proposes, head disposes — see `probe.json`.

## Review split (gliner2_review.jsonl, 2026-09-06)

421 passages / 473 spans from the annotator boundary set (190) and JDC
operational review (283), added as origins `annotator190` / `jdc283` with
`split: review`. Labels are the reconciled v2.4 gold (3 fresh single-span
runs + Sonnet tiebreak + human adjudication; see
`rafmacalaba/datause-v24-sample-verdicts`, doctrine_findings.md). Span
entries carry `key`, `review_tier`, and char offsets.