Instructions to use rafmacalaba/gliner_datause_probe with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- GLiNER
How to use rafmacalaba/gliner_datause_probe with GLiNER:
from gliner import GLiNER model = GLiNER.from_pretrained("rafmacalaba/gliner_datause_probe") - Notebooks
- Google Colab
- Kaggle
gliner_datause_probe
Frozen-encoder probe of rafmacalaba/gliner_datause (proposals: NAMED_DATA, DESCRIPTIVE_DATA, VAGUE_DATA). The Luna v2.4
keep/drop boundary (v2.4-patched-2026-09-05) is read off the frozen representation with a
small MLP head on [start; end; mean; 卤64-token window] span features: extractor proposes, head disposes.
Results (holdout)
head AUROC: 0.8169
holdout-190 (annotator): head AUROC 0.8889 路 best-F1 thr 0.4: prec 0.8099 路 rec 0.8909 路 f1 0.8485
holdout-jdc283: head AUROC 0.5943 路 best-F1 thr 0.3: prec 0.8514 路 rec 0.3073 路 f1 0.4516
holdout-human473 (190+283 combined): head AUROC 0.6254 路 best-F1 thr 0.4: prec 0.8290 路 rec 0.5079 路 f1 0.6299
raw GLiNER score AUROC: 0.5716
raw GLiNER score AUROC: 0.5716
best-F1 thr 0.4: precision 0.8143 路 recall 0.7674 路 f1 0.7901
residual-leak diagnostic: 50 high-confidence raw-score FPs, head expelled 40 (80.0%)
Labels come from the Luna v2.4 review (probe_reviewed config of
rafmacalaba/datause-displacement-reviewed, prompt v2.4-patched-2026-09-05); holdout
includes the 190 annotator + 283 JDC human-adjudicated spans. Head: head.pt;
per-span predictions: holdout_predictions.jsonl (key, v2.4 label,
head_score, raw_score).
- Downloads last month
- -