gliner-datause-displacement

Fine-tune of urchade/gliner_large-v2.1 for data-use mention extraction with a single DATA_MENTION class, trained on rafmacalaba/data-use-mentions-tiered — the tiered copy of rafmacalaba/data-use-mentions where Luna/classifier-judged T3 (non-mention) and junk spans are untagged hard negatives (text stays, span removed). The extractor owns the mention boundary only (T1 evidential ∪ T2 declaration vs T3/junk); specificity detail is recovered downstream by the multitask SFT model.

Labels

• DATA_MENTION — a real data mention that carries an analytic or declarative use (T1 evidential ∪ T2 declaration)

Training

• base model: urchade/gliner_large-v2.1
• dataset: rafmacalaba/datause-displacement-reviewed (gliner_reviewed config)
• epochs: 8
• learning rate: 5e-06
• batch size: 16
• precision: bf16
• checkpoint selection: val span-F0.5 (post-hoc sweep of epoch checkpoints; eval_loss was explicitly not used)

Evaluation (tiered holdout)

Gold = T1∪T2 spans; a true-FP cluster matching a dropped T3/junk span counts as a T3 leak (lower is better). Label-agnostic Hungarian matching, jaccard >= 0.5 — identical to prior data-use-mentions evals.

thr	tp	fp	fn	precision	recall	f0.5	f1	t3_leak	t3_leak%
0.10	1232	2118	35	0.3678	0.9724	0.4200	0.5337	439	20.7%
0.20	1205	1607	62	0.4285	0.9511	0.4814	0.5908	371	23.1%
0.30	1175	1284	92	0.4778	0.9274	0.5291	0.6307	328	25.6%
0.40	1140	1031	127	0.5251	0.8998	0.5728	0.6632	283	27.5%
0.50	1069	772	198	0.5807	0.8437	0.6193	0.6879	242	31.4%
0.60	963	531	304	0.6446	0.7601	0.6648	0.6976	184	34.6%
0.70	779	286	488	0.7315	0.6148	0.7047	0.6681	110	38.5%

Best F0.5: 0.7047 (thr=0.7) Best F1: 0.6976 (thr=0.6)

Full per-doc predictions (raw scores, gold spans with tier decisions): holdout_predictions.jsonl on this repo.

Corpus breakdown (holdout, best F0.5)

corpus	examples	spans	thr	precision	recall	f0.5	f1
prwp	0	0	0.10	1.0000	1.0000	1.0000	1.0000
fcv	0	0	0.10	1.0000	1.0000	1.0000	1.0000