⚠️ Experimental checkpoint — negative result (private)
Treatment arm of a head-initialization A/B (GLiNER2 working paper, §10.7): mmBERT-base fine-tuned on WikiEvents from the broad combined base
whr778/mmbert-base-combined. On the WikiEvents blind test it gave no reliable lift over the RAMS-only-base control: argument-strict F1 at the floor (0.007 vs 0.005), trigger edge is precision-only within noise (0.133 vs 0.085), and event-type regressed (0.573 vs 0.944). Not for production; kept for reproducibility.
mmbert_base_wikievents_combined
A GLiNER2 multi-task information-extraction model (entities, relations, events, and classification) fine-tuned from ./out/mmbert-base-combined/best.
⚠️ 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:
./out/mmbert-base-combined/best - Library:
gliner2 - Tasks: entity, relation, event, and classification extraction
- Experiment:
mmbert_base_wikievents_combined
Training data
1 dataset used for this run. 206 training records (val: 20, test: 20).
| Dataset | Task(s) | Train | Val | Test | Language | License | Source |
|---|---|---|---|---|---|---|---|
| WikiEvents | NER + event extraction | 206 | 20 | 20 | en | see source | link |
Dataset notes
- WikiEvents — KAIROS-ontology event extraction co-trained with typed entity mentions; 49 event types, 57 argument roles.
Training procedure
| Setting | Value |
|---|---|
| Trained on | 2026-08-03 |
| Duration | 10m 11s |
| Throughput | 4.4 samples/s |
| Epochs | 15 |
| Batch size | 2 (× 16 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_event_argument_strict_micro_f1 |
| Seed | 42 |
| Architecture | 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 | 0.196 → 0.249 | 0.472 → 0.599 | 0.277 → 0.352 | 1602 |
| event_type | 1.000 → 1.000 | 0.402 → 0.402 | 0.573 → 0.573 | 122 |
| event_trigger | 0.217 → 0.217 | 0.096 → 0.096 | 0.133 → 0.133 | 239 |
| event_argument | 0.023 → 0.264 | 0.004 → 0.048 | 0.007 → 0.081 | 515 |
| event | 0.306 → 0.393 | 0.084 → 0.113 | 0.132 → 0.176 | 876 |
Best checkpoint (validation)
Micro precision / recall / F1, strict → relaxed.
| Category | Precision | Recall | F1 | Support |
|---|---|---|---|---|
| entity | 0.293 → 0.391 | 0.209 → 0.278 | 0.244 → 0.325 | 1427 |
| event_type | 1.000 → 1.000 | 0.140 → 0.140 | 0.245 → 0.245 | 129 |
| event_trigger | 0.400 → 0.400 | 0.030 → 0.030 | 0.055 → 0.055 | 269 |
| event_argument | 0.429 → 0.714 | 0.007 → 0.013 | 0.014 → 0.026 | 416 |
| event | 0.644 → 0.689 | 0.036 → 0.040 | 0.068 → 0.075 | 814 |
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: ./out/mmbert-base-combined/best — see model card
Unverified — verify the upstream terms before redistribution
- ./out/mmbert-base-combined/best (see model card)
- WikiEvents (see source)
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).