gliner_datause

Fine-tune of urchade/gliner_large-v2.1 for data-use mention extraction (dataset / survey / census / registry mentions in economics research papers).

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

Training

• base model: urchade/gliner_large-v2.1
• dataset: rafmacalaba/data-use-mentions (gliner config)
• corpus: all
• epochs: 5
• learning rate: 5e-06
• batch size: 8
• precision: bf16

Evaluation (holdout)

thr	tp	fp	fn	precision	recall	f0.5	f1
0.10	16046	9301	123	0.6331	0.9924	0.6825	0.7730
0.20	16016	6724	153	0.7043	0.9905	0.7475	0.8233
0.30	15971	5288	198	0.7513	0.9878	0.7890	0.8534
0.40	15895	4053	274	0.7968	0.9831	0.8282	0.8802
0.50	15694	2871	475	0.8454	0.9706	0.8678	0.9037
0.60	14185	1753	1984	0.8900	0.8773	0.8874	0.8836
0.70	11161	827	5008	0.9310	0.6903	0.8703	0.7928

Best F0.5: 0.8874 (thr=0.6) Best F1: 0.9037 (thr=0.5)

Evaluation breakdown (holdout)

group	examples	spans	thr	precision	recall	f0.5	f1
overall	12531	16417	0.60	0.8900	0.8773	0.8874	0.8836
prwp	9079	12380	0.60	0.8886	0.8935	0.8896	0.8911
fcv	3452	4037	0.50	0.8670	1.0000	0.8907	0.9287
general_prwp	9079	12380	0.60	0.8886	0.8935	0.8896	0.8911
fcv_pads_east_asia	784	863	0.50	0.8877	1.0000	0.9081	0.9405
jdc_operational	163	178	0.60	0.9193	0.8555	0.9058	0.8862
refugee_pads	803	875	0.50	0.8475	1.0000	0.8742	0.9175