gliner-datause-mentions-catch-all

GLiNER catch-all proposer (catch_all) trained on rafmacalaba/data-use-mentions (config catch_all, column ner):

• arm catch_all: ner column (every candidate span is DATA_MENTION: keeps and drops alike) — recall-oriented proposer for the probe cascade. The probe head owns keep/drop. Eval: gold = DATA_MENTION spans; predictions = DATA_MENTION predictions only.

Training

• base model: urchade/gliner_large-v2.1
• dataset: rafmacalaba/data-use-mentions (config catch_all, column ner)
• labels: DATA_MENTION
• corpus: all
• epochs: 5
• learning rate: 5e-06
• batch size: 16
• precision: bf16

Evaluation (holdout, DATA_MENTION only)

thr	tp	fp	fn	precision	recall	f0.5	f1
0.10	21326	10740	364	0.6651	0.9832	0.7111	0.7934
0.20	21168	8164	522	0.7217	0.9759	0.7613	0.8298
0.30	20987	6705	703	0.7579	0.9676	0.7922	0.8500
0.40	20718	5369	972	0.7942	0.9552	0.8219	0.8673
0.50	20227	4072	1463	0.8324	0.9325	0.8507	0.8796
0.60	19318	2901	2372	0.8694	0.8906	0.8736	0.8799
0.70	17611	1869	4079	0.9041	0.8119	0.8840	0.8555

Best F0.5: 0.8840 (thr=0.7) Best F1: 0.8799 (thr=0.6)

Breakdown (best-F0.5 operating point)

group	examples	spans	thr	precision	recall	f0.5	f1
overall	18709	22052	0.70	0.9041	0.8119	0.8840	0.8555
prwp	9188	12359	0.70	0.8981	0.7626	0.8672	0.8248
fcv	9521	9693	0.70	0.9110	0.8766	0.9039	0.8935
fcv_pads_east_africa	6794	6784	0.70	0.9210	0.8789	0.9123	0.8995
general_prwp	9188	12359	0.70	0.8981	0.7626	0.8672	0.8248
jdc_operational	158	204	0.60	0.9624	0.9133	0.9521	0.9372
refugee_pads	687	743	0.70	0.8989	0.8852	0.8961	0.8920
reliefweb	1882	1962	0.70	0.8771	0.8661	0.8749	0.8716