gliner-datause-catchall-singlepass
Single-pass catch-all cascade in one bundle: the rafmacalaba/gliner-datause-mentions-catch-all encoder (frozen, byte-identical) plus an inference-native probe head (probe_head.pt, trained by outputs/gliner-datause-catchall-infer-probe with --feature-source infer). One forward per doc yields proposals (GLiNER DATA_MENTION @ 0.1) and keep/drop (probe_score) together — see training/singlepass_infer.py.
- head: input_dim 2048, hidden 256, context_radius 64
- keep knob: probe_score >= thr (global best 0.5 @ F1 0.8467)
Per-origin best-F1 thresholds
| origin | thr | f1 |
|---|---|---|
fcv_pads_east_africa |
0.5 | 0.8046 |
general_prwp |
0.4 | 0.8796 |
jad_paddy_docs |
0.1 | 0.9693 |
jdc_operational |
0.6 | 0.8019 |
refugee_pads |
0.5 | 0.8387 |
reliefweb |
0.4 | 0.7738 |
Inference
from training.singlepass_infer import load_bundle, predict_keep_drop
model, head, bundle = load_bundle('rafmacalaba/gliner-datause-catchall-singlepass', 'rafmacalaba/gliner-datause-catchall-singlepass', 'cuda')
rows = predict_keep_drop(texts, model, head, bundle,
propose_thr=0.1, keep_thr=0.3)
Files: pytorch_model.bin + gliner_config.json (encoder),
probe_head.pt + head_config.json (head), thresholds.json
(operating points), holdout_metrics.json (sweep).