---
task_categories:
  - token-classification
tags:
  - ner
  - span-extraction
  - data-mentions
license: cc-by-4.0
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
---

# datause-agents-v3 — data-use mentions

## Annotation workflow

Labels are agent-produced under the project data-use doctrine; they are not owner rulings. The training and holdout splits are unchanged.

### Validation review refresh

The `val` labels in both `gliner` and `gliner2` were refreshed on 2026-09-29T17:54:34+00:00 using the `gliner_val_review` passage-review lane and shards 01–04. The review covered 1,260 passages and 2,171 candidate mentions drawn from published validation labels plus GLiNER predictions scoring at least 0.6. All decisions remain `annotated_by: agent`, `owner_reviewed: false`. The other 862 candidate-free validation passages retain their existing empty labels. Because review scope is candidate-generated and does not assess mentions missed by the candidate policy, this split is an agent-reviewed audit set, not independent full-recall gold.

The review-specific source metadata and span counts are in `split_stats.json`.

## Labels

- `NAMED_DATA` — a proper name, title, or acronym of a specific data source
- `DESCRIPTIVE_DATA` — a source described in words but not named
- `VAGUE_DATA` — generic data wording with no identifiable source

## Provenance

- bundle: `agent_annotations_v3` (agent-annotated; `owner_reviewed: false`)
- lanes: origin_strat_1, origin_strat_2, origin_strat_3, umar_pads, umar_pads_500
- carve: `agents-v3` (origin-stratified) — document-disjoint, so no
  two passages of one document straddle train and eval, and stratified by origin, so each
  split holds each corpus family's exact share
- guard: 77 eval-text and 253 duplicate-text passages excluded;
  128 eval documents checked, 0 leaks
- **not evaluable**: jdc_operational have no documents in val/holdout (too few to stratify), so they are train-only
  and no split metric says anything about them
- labels: `labels.json`, per-split counts and per-origin document counts: `split_stats.json`

## Configs

- `gliner` — `{"tokenized_text": [...], "ner": [[start, end, LABEL], ...]}` (`ner` is inclusive on both ends)
- `gliner2` — `{"input": "...", "output": {"entities": {grade_key: [span strings]}}}`

## Splits

21,073 passages · 16,630 spans · 12,831 documents, document-disjoint.

| split | rows | spans | documents | NAMED | DESCRIPTIVE | VAGUE |
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
| train | 16,843 | 12,891 | 10,265 | 9,723 | 3,142 | 26 |
| val | 2,122 | 1,821 | 1,283 | 1,275 | 538 | 8 |
| holdout | 2,108 | 1,918 | 1,283 | 1,360 | 548 | 10 |

## Usage

```bash
uv run python training/finetune_gliner.py \
    --dataset rafmacalaba/datause-agents-v3 --config gliner \
    --epochs 5 --batch-size 16 --output-dir models/gliner_agents_v3
```

```python
from datasets import load_dataset
ds = load_dataset("rafmacalaba/datause-agents-v3", "gliner")   # ds['train'] / ds['val'] / ds['holdout']
```