whr778/mmbert-base-combined

⚠️ Experimental checkpoint — negative result (private)

Research artifact from a head-initialization A/B (GLiNER2 working paper, §10.7). This broad combined base warms the GLiNER2 heads for only 2 epochs (eval_loss checkpoint selection) on a mix of synthetic multi-task data, GLiNER multilingual/multi-task NER, and RAMS events. It exists only as the base for a downstream WikiEvents fine-tune; that fine-tune showed no reliable improvement over the RAMS-only base. Its own RAMS argument-strict F1 is weak (0.028; the RAMS-only mmBERT base is 0.050 — both cold-start). Not for production; kept for reproducibility.

mmbert_base_combined

A GLiNER2 multi-task information-extraction model (entities, relations, events, and classification) fine-tuned from jhu-clsp/mmBERT-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: jhu-clsp/mmBERT-base
• Library: gliner2
• Tasks: entity, relation, event, and classification extraction
• Experiment: mmbert_base_combined

Training data

4 datasets used for this run. 96,404 training records (val: 11,989, test: 12,102).

Dataset	Task(s)	Train	Val	Test	Language	License	Source
⚠️ synthetic_sonnet5_1k	unknown	—	—	—	—	UNKNOWN — not in registry	—
GLiNER multilingual synthetic	NER (multilingual)	77,259	9,598	9,749	en, de, fr, es, it, pl, nl, pt	see card	link
GLiNER multi-task synthetic	NER (multi-task)	10,319	1,276	1,288	en	Apache-2.0	link
RAMS	Event extraction (trigger + args)	7,329	924	871	en	see source	link

Dataset notes

• GLiNER multilingual synthetic — Multilingual synthetic NER (German, French, Polish, and others) for multilingual encoder training.
• GLiNER multi-task synthetic — Dense multi-type synthetic NER (~10 types per record), open vocabulary.
• RAMS — Multi-sentence event extraction with triggers and typed arguments; 139 event types, 65 argument roles.

Training procedure

Setting	Value
Trained on	2026-08-03
Duration	1h 40m
Throughput	31.8 samples/s
Epochs	2
Batch size	4 (× 8 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=8192, struct_loss=bce_posweight, struct_pos_weight=4.0

Evaluation

Decision threshold: 0.5 (calibrated against the validation set).

Blind test (held-out test splits)

Micro precision / recall / F1, strict → relaxed.

Category	Precision	Recall	F1	Support
entity	0.351 → 0.433	0.255 → 0.315	0.295 → 0.364	61530
relation	0.001 → 0.022	0.001 → 0.016	0.001 → 0.018	1512
classification	0.385 → 0.433	0.294 → 0.331	0.333 → 0.375	762
event_type	1.000 → 1.000	0.791 → 0.791	0.883 → 0.883	1657
event_trigger	0.275 → 0.280	0.384 → 0.391	0.321 → 0.326	1729
event_argument	0.026 → 0.214	0.032 → 0.269	0.028 → 0.238	4970
event	0.215 → 0.334	0.255 → 0.398	0.234 → 0.363	8356

Best checkpoint (validation)

Micro precision / recall / F1, strict → relaxed.

Category	Precision	Recall	F1	Support
entity	0.346 → 0.423	0.248 → 0.303	0.289 → 0.353	62301
relation	0.001 → 0.033	0.001 → 0.024	0.001 → 0.028	1571
classification	0.410 → 0.450	0.307 → 0.337	0.351 → 0.385	766
event_type	1.000 → 1.000	0.778 → 0.778	0.875 → 0.875	1695
event_trigger	0.268 → 0.272	0.359 → 0.365	0.307 → 0.312	1777
event_argument	0.021 → 0.219	0.024 → 0.260	0.022 → 0.238	5204
event	0.212 → 0.337	0.240 → 0.384	0.225 → 0.359	8676

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: mmBERT-base — see model card

Unverified — verify the upstream terms before redistribution

• GLiNER multilingual synthetic (see card)
• RAMS (see source)
• mmBERT-base (see model card)
• synthetic_sonnet5_1k (unknown) (unspecified)

Permissive

• GLiNER multi-task synthetic (Apache-2.0)

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).