whr778/deberta-base-fromenc-synthetic

⚠️ From-scratch negative result (public)

Fresh GLiNER2 heads on the raw microsoft/deberta-v3-base encoder (from_encoder), trained on synthetic_sonnet5_1k only (15 epochs). This tests whether ~1,500 synthetic records can teach the extraction tasks from scratch. They cannot — every span/relation task collapses. Blind test (synthetic held-out split), strict micro-F1, vs the warm-start counterpart whr778/gliner2-base-v1-synthetic:

Task	From-encoder (this)	Warm-start (base-v1)
Entity	0.141	0.904
Relation	0.000	0.657
Event type	0.998	0.956
Event trigger	0.221	0.838
Event argument	0.000 (0.168 relaxed)	0.702
Classification	0.356	0.835

Only coarse event-type (few classes) is learnable from scratch; fine-grained span extraction and relations need either a warm start or the ~10⁵–10⁶-scale IE curriculum the fastino heads saw. The synthetic corpus is good (the warm-start model works) but is an adaptation set, not a from-scratch pretraining set. Not for use; documented negative result.

deberta_base_fromenc_synthetic

A GLiNER2 multi-task information-extraction model (entities, relations, events, and classification) fine-tuned from microsoft/deberta-v3-base.

⚠️ License at a glance

• Effective license: Unverified — review required
• Commercial use: Unverified
• All dataset licenses verified: No

See License for the full determination and per-dataset terms.

Model details

• Base model: microsoft/deberta-v3-base
• Library: gliner2
• Tasks: entity, relation, event, and classification extraction
• Experiment: deberta_base_fromenc_synthetic

Training data

1 dataset used for this run. 1,497 training records (val: 191, test: 194).

Dataset	Task(s)	Train	Val	Test	Language	License	Source
⚠️ synthetic_sonnet5_1k	unknown	—	—	—	—	UNKNOWN — not in registry	—

Training procedure

Setting	Value
Trained on	2026-08-03
Duration	40m 38s
Throughput	9.0 samples/s
Epochs	15
Batch size	8 (× 4 grad-accum)
Encoder LR	2e-05
Task-head LR	0.0005
Weight decay	0.01
Scheduler	cosine_restarts (warmup 0.05)
Precision	bf16
Max grad norm	1.0
Best-checkpoint metric	eval_loss
Seed	42
Architecture	max_width=20, max_len=512, struct_loss=bce_posweight, struct_pos_weight=4.0

Evaluation

Decision threshold: 0.3 (calibrated against the validation set).

Blind test (held-out test splits)

Micro precision / recall / F1, strict → relaxed.

Category	Precision	Recall	F1	Support
entity	0.082 → 0.099	0.512 → 0.616	0.141 → 0.170	6010
relation	0.000 → 0.024	0.000 → 0.019	0.000 → 0.021	1512
classification	0.411 → 0.565	0.314 → 0.432	0.356 → 0.490	762
event_type	1.000 → 1.000	0.996 → 0.996	0.998 → 0.998	809
event_trigger	0.130 → 0.133	0.725 → 0.742	0.221 → 0.226	881
event_argument	0.000 → 0.116	0.000 → 0.303	0.000 → 0.168	2954
event	0.109 → 0.176	0.311 → 0.509	0.161 → 0.262	4644

Best checkpoint (validation)

Micro precision / recall / F1, strict → relaxed.

Category	Precision	Recall	F1	Support
entity	0.088 → 0.115	0.078 → 0.102	0.083 → 0.108	5900
relation	0.000 → 0.042	0.000 → 0.028	0.000 → 0.033	1571
classification	0.394 → 0.555	0.295 → 0.415	0.338 → 0.475	766
event_type	1.000 → 1.000	0.985 → 0.985	0.992 → 0.992	799
event_trigger	0.162 → 0.164	0.594 → 0.600	0.254 → 0.257	881
event_argument	0.002 → 0.164	0.001 → 0.096	0.001 → 0.121	3022
event	0.228 → 0.278	0.279 → 0.345	0.251 → 0.308	4702

License

Effective license: Unverified — review required. This model is a derivative of its base model and every training dataset, so the most restrictive term across all of them governs the whole model.

• Commercial use: Unverified
• Share-alike obligation: No
• All licenses verified: No
• Base model: DeBERTa-v3-base — mit

Unverified — verify the upstream terms before redistribution

• synthetic_sonnet5_1k (unknown) (unspecified)

Permissive

• DeBERTa-v3-base (mit)

License strings are copied verbatim from each dataset's card/source and from tools/train/dataset_registry.yaml. "see card"/"see source"/"other" mean the upstream declares no clear license — treat as unverified. This summary is informational, not legal advice; confirm terms before redistribution or commercial use.

Citation

If you use this model, please cite GLiNER2 and the underlying datasets (linked in Training data).

---

Model card generated automatically at the end of training (2026-08-03).