Research checkpoint from the mmBERT head-init data-scaling curve (private). The warmed base whr778/scaling-mmbert-100k (mmBERT-base from_encoder, warmed on ~100,000 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).

This is the N=100k 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.

Quantifies whether warming mmBERT's heads on ~100,000 records lifts downstream RAMS argument extraction. See SCALING_CURVE_EXPERIMENT.md and PAPER.md §10.

Research artifact, private. One point on a scaling curve, not a general release.