gliner-datause-displacement-withnonmention

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_nm config)
• epochs: 5
• 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	254	440	66	0.3660	0.7937	0.4102	0.5010	47	10.7%
0.20	250	351	70	0.4160	0.7812	0.4589	0.5429	45	12.8%
0.30	242	289	78	0.4557	0.7562	0.4951	0.5687	44	15.2%
0.40	233	226	87	0.5076	0.7281	0.5404	0.5982	40	17.7%
0.50	217	177	103	0.5508	0.6781	0.5723	0.6078	34	19.2%
0.60	177	89	143	0.6654	0.5531	0.6395	0.6041	23	25.8%
0.70	125	39	195	0.7622	0.3906	0.6404	0.5165	16	41.0%

Best F0.5: 0.6404 (thr=0.7) Best F1: 0.6078 (thr=0.5)

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