whr778/gliner2-base-v1-synthetic

✅ Sanity-check checkpoint (private)

Fine-tune of fastino/gliner2-base-v1 on our multi-task synthetic corpus (synthetic_sonnet5_1k) only, to check whether our synthetic data trains cleanly across all five GLiNER2 tasks. It does — blind test on the synthetic held-out split (in-distribution):

Task	Strict F1	Relaxed F1	Support
Entity	0.904	0.933	6,010
Relation	0.657	0.745	1,512
Event type	0.956	—	809
Event trigger	0.838	—	881
Event argument	0.702	0.894	2,954
Classification	0.835	0.841	762

These are on the synthetic test split (in-distribution): they show the data is coherent and learnable, not a claim about the original public benchmarks. Warm-started (from_pretrained), 10 epochs, eval_loss selection (best = epoch 7). Experimental; kept for reproducibility.

gliner2_base_v1_synthetic

A GLiNER2 multi-task information-extraction model (entities, relations, events, and classification) fine-tuned from fastino/gliner2-base-v1.

⚠️ 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: fastino/gliner2-base-v1
• Library: gliner2
• Tasks: entity, relation, event, and classification extraction
• Experiment: gliner2_base_v1_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	32m 55s
Throughput	7.4 samples/s
Epochs	10
Batch size	8 (× 4 grad-accum)
Encoder LR	1e-05
Task-head LR	0.0003
Weight decay	0.01
Scheduler	cosine_restarts (warmup 0.05)
Precision	bf16
Max grad norm	1.0
Best-checkpoint metric	eval_loss
Seed	42

Evaluation

Decision threshold: 0.7 (calibrated against the validation set).

Blind test (held-out test splits)

Micro precision / recall / F1, strict → relaxed.

Category	Precision	Recall	F1	Support
entity	0.914 → 0.943	0.894 → 0.923	0.904 → 0.933	6010
relation	0.790 → 0.895	0.563 → 0.638	0.657 → 0.745	1512
classification	0.964 → 0.971	0.736 → 0.741	0.835 → 0.841	762
event_type	1.000 → 1.000	0.916 → 0.916	0.956 → 0.956	809
event_trigger	0.847 → 0.866	0.830 → 0.848	0.838 → 0.857	881
event_argument	0.737 → 0.940	0.670 → 0.851	0.702 → 0.893	2954
event	0.805 → 0.935	0.743 → 0.862	0.773 → 0.897	4644

Best checkpoint (validation)

Micro precision / recall / F1, strict → relaxed.

Category	Precision	Recall	F1	Support
entity	0.886 → 0.917	0.919 → 0.952	0.902 → 0.934	5900
relation	0.716 → 0.837	0.604 → 0.706	0.655 → 0.766	1571
classification	0.953 → 0.960	0.713 → 0.718	0.816 → 0.822	766
event_type	1.000 → 1.000	0.962 → 0.962	0.981 → 0.981	799
event_trigger	0.796 → 0.807	0.875 → 0.888	0.834 → 0.845	881
event_argument	0.649 → 0.893	0.666 → 0.911	0.657 → 0.902	3022
event	0.734 → 0.893	0.755 → 0.915	0.745 → 0.904	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: gliner2-base-v1 — see model card

Unverified — verify the upstream terms before redistribution

• gliner2-base-v1 (see model card)
• synthetic_sonnet5_1k (unknown) (unspecified)

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).