---
task_categories:
  - token-classification
tags:
  - ner
  - span-extraction
  - data-mentions
license: apache-2.0
configs:
  - config_name: gliner
    data_files:
      - split: train
        path: gliner_train.jsonl
      - split: val
        path: gliner_val.jsonl
      - split: holdout
        path: gliner_holdout.jsonl
  - config_name: gliner2
    data_files:
      - split: train
        path: gliner2_train.jsonl
      - split: val
        path: gliner2_val.jsonl
      - split: holdout
        path: gliner2_holdout.jsonl
  - config_name: probe_splits
    data_files:
      - split: train
        path: probe_train.jsonl
      - split: val
        path: probe_val.jsonl
      - split: holdout
        path: probe_holdout.jsonl
  - config_name: probe_candidates
    data_files:
      - split: pool
        path: probe_candidate_pool.jsonl
---

# Datause NER (catch-all DATA_MENTION + probe configs)

Catch-all NER views over `rafmacalaba/datause-probe-v3` passages (29,346 spans grouped into passage examples). Single entity type **DATA_MENTION**: every candidate span is tagged, keeps and drops alike — the probe head (not NER tags) owns the keep/drop boundary. No NAMED/DESCRIPTIVE/VAGUE subtypes, no NON_MENTION.

## Per-origin thresholds (head best-F1, published holdout sweep)

| origin | threshold |
|---|---|
| `fcv_pads_east_africa` | 0.5 |
| `general_prwp` | 0.3 |
| `jad_paddy_docs` | 0.3 |
| `jdc_operational` | 0.5 |
| `refugee_pads` | 0.6 |
| `reliefweb` | 0.3 |

## Columns

- `gliner`: `tokenized_text` + `ner` — every candidate span tagged `DATA_MENTION` (catch-all; keeps and drops alike).
- `gliner2`: `input` + `output` (`data_mention` = all span strings).
- Keep/drop lives in `spans[].luna_label` + `head_score`/`threshold` traceability, not in tags — the probe head owns that boundary.
- Traceability on every row: `corpus_id`, `page`, `chunk`, `split`, `spans` (per-span text, pred, luna_label, head_score, threshold, char offsets, key, source). Luna keep/drop verdicts are preserved in `spans[].luna_label` for audit, but tagging follows the head-threshold operating point.

## Splits

Doc-disjoint `train` / `val` / `holdout`, inherited from `rafmacalaba/datause-probe-v3` (`outputs/probe_v3_data`), so NER splits match probe splits. See `split_stats.json`.

## Usage

```python
from datasets import load_dataset
gliner  = load_dataset("rafmacalaba/datause-ner", "gliner")
gliner2 = load_dataset("rafmacalaba/datause-ner", "gliner2")
# catch-all supervision: every span is DATA_MENTION:
#   gliner['train']['ner'] / gliner2['train']['output']
```

## Files

- `gliner_{train,val,holdout}.jsonl`, `gliner2_{train,val,holdout}.jsonl`
- `labels.json`, `thresholds.json`, `split_stats.json`
- `probe_{train,val,holdout}.jsonl`, `probe_candidate_pool.jsonl`, `probe_split_stats.json` — mirror of `rafmacalaba/datause-probe-v3` (same data, `probe_*` configs)

## Probe configs (same data as `rafmacalaba/datause-probe-v3`)

```python
splits = load_dataset("rafmacalaba/datause-ner", "probe_splits")
cands  = load_dataset("rafmacalaba/datause-ner", "probe_candidates")
```