gliner_datause_extended
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 sourceDESCRIPTIVE_DATA — a source described in words but not namedVAGUE_DATA — generic data wording with no identifiable source
Training
- base model:
urchade/gliner_large-v2.1
- dataset:
rafmacalaba/data-use-mentions-extended (gliner config)
- epochs: 5
- learning rate: 5e-06
- batch size: 16
- precision: bf16
Evaluation (holdout)
| thr | tp | fp | fn | precision | recall | f0.5 | f1 |
|---|
| 0.10 | 12283 | 8280 | 281 | 0.5973 | 0.9776 | 0.6477 | 0.7416 |
| 0.20 | 12192 | 6065 | 372 | 0.6678 | 0.9704 | 0.7122 | 0.7911 |
| 0.30 | 12062 | 4883 | 502 | 0.7118 | 0.9600 | 0.7506 | 0.8175 |
| 0.40 | 11845 | 3857 | 719 | 0.7544 | 0.9428 | 0.7858 | 0.8381 |
| 0.50 | 11498 | 2837 | 1066 | 0.8021 | 0.9152 | 0.8224 | 0.8549 |
| 0.60 | 10519 | 1798 | 2045 | 0.8540 | 0.8372 | 0.8506 | 0.8455 |
| 0.70 | 8328 | 892 | 4236 | 0.9033 | 0.6628 | 0.8422 | 0.7646 |
Best F0.5: 0.8506 (thr=0.6)
Best F1: 0.8549 (thr=0.5)
NER holdout comparison
device: NVIDIA H100 NVL
rafmacalaba/data-use-mentions-extended (n=9249)
| model | backend | best F0.5 | thr | best F1 | thr | wall-clock (s) | texts/s |
|---|
rafmacalaba/gliner_datause_extended | gliner | 0.8506 | 0.6 | 0.8549 | 0.5 | 202.7 | 45.6 |
ai4data/gliner2_datause | gliner2 | 0.8634 | 0.7 | 0.8624 | 0.6 | 180.8 | 51.1 |
F0.5 by threshold (sweet spots side-by-side):
| thr | rafmacalaba/gliner_datause_extended | ai4data/gliner2_datause |
|---|
| 0.1 | 0.6476 | 0.7321 |
| 0.2 | 0.7121 | 0.7712 |
| 0.3 | 0.7505 | 0.7979 |
| 0.4 | 0.7858 | 0.8201 |
| 0.5 | 0.8224 | 0.8363 |
| 0.6 | 0.8506 | 0.8523 |
| 0.7 | 0.8422 | 0.8634 |
FCV cross-corpus validation (precision)
Validated on the World Bank Fragility, Conflict & Violence (FCV) document corpus with
the gpt-5.6-luna judge (Luna validation). Precision-only: there is no gold set of all
true mentions, so recall / F1 are not reported here.
Precision by config × label
| config | NAMED_DATA | DESCRIPTIVE_DATA | VAGUE_DATA | overall |
|---|
| fcv_pads_east_asia | 80.1% | 81.6% | 91.8% | 81.9% |
| jdc_operational | 86.8% | 91.3% | 91.5% | 89.1% |
| refugee_pads | 83.8% | 92.8% | 91.9% | 89.1% |
| reliefweb | 83.7% | 85.4% | 95.5% | 85.6% |
Specificity agreement (extractor label == judge specificity)
| config | NAMED_DATA | DESCRIPTIVE_DATA | VAGUE_DATA |
|---|
| fcv_pads_east_asia | 82.8% | 88.3% | 82.1% |
| jdc_operational | 90.2% | 83.2% | 71.7% |
| refugee_pads | 89.4% | 86.5% | 77.5% |
| reliefweb | 92.6% | 82.5% | 73.4% |
Notes:
VAGUE_DATA is the safest label (91–95% precision); NAMED_DATA is the weakest (80–87%).- The extractor under-specifies:
VAGUE_DATA is often judged descriptive, and
DESCRIPTIVE_DATA → named leaks occur.
reliefweb carries more bad-JSON judge errors (OCR/HTML artifacts) than the other configs.
