Downloads · 30 days
9
35% of all-time downloads
scaling-mmbert-40k-rams
whr778/scaling-mmbert-40k-rams
scaling-mmbert-40k-rams is a token classification model from whr778. Use it when you need labels on individual words, such as names. It is set up for gliner2.
Research checkpoint from the mmBERT head-init data-scaling curve (private). The warmed base whr778/scaling-mmbert-40k (mmBERT-base fromencoder, warmed on ~40,000 structure/argument records) fine-tuned on RAMS under th…
All-time downloads
26
Public
Parameters
337M
1.4 GB on disk
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors1.3 GB · 98%
At a glance
- Task
- Token Classification
- Library
- gliner2
- Model type
- extractor
- Access
- Public
- Created
- Aug 4, 2026
- Updated
- Aug 4, 2026
- SHA
- 2b5973e7
Model card
From the Hugging Face model README
scaling-mmbert-40k-rams (Stage B, RAMS fine-tune)
Research checkpoint from the mmBERT head-init data-scaling curve (private).
The warmed base whr778/scaling-mmbert-40k (mmBERT-base
from_encoder, warmed on ~40,000 structure/argument records) fine-tuned on RAMS
under the fixed mmbert-base-rams recipe (15 epochs, bce_posweight 4.0, native
long-context, argument-strict checkpoint selection).
RAMS blind-test (871 docs), strict micro-F1
| metric | score |
|---|---|
| event_argument_strict | 0.115 |
| event_trigger_strict | 0.706 |
This is the N=40k point's y-value on the curve: RAMS argument-strict F1 vs Stage-A corpus size. Curve context: N=0 (fresh heads → RAMS) = 0.050; a DeBERTa-v3 fastino warm-start (254K) reaches 0.462 as a cross-encoder reference.
Purpose
Quantifies whether warming mmBERT's heads on ~40,000 records lifts downstream RAMS
argument extraction. See SCALING_CURVE_EXPERIMENT.md and PAPER.md §10.
Caveats
Research artifact, private. One point on a scaling curve, not a general release.
- Task
- Token Classification
- Library
- gliner2
- Type
- extractor
- Languages
- multilingual
- Created
- Aug 4, 2026
- Updated
- Aug 4, 2026