Modelpublic

Montalte/qwen4b-code-nothink-localize

sourceHugging Faceapache-2.0updated 13d agoView on Hugging Face

0likes231downloads

Model Card

Unified Qwen3-4B merge artifact for directional math↔code transfer experiments.

Field	Value
Base	`Qwen/Qwen3-4B-Base@906bfd4b4dc7f14ee4320094d8b41684abff8539`
Source specialist	`modrill/code-nothink-q4b-20260908`
Domain	code
Mode	nothink
Method	localize

localize = Plan B Localize-and-Stitch on source-only MergeBench validation.

—sparsity (keep) target: 0.1
—lr=1e7, epochs=10, n-shot=64, seed=42, mask-profile=plan_b
—task=coding
—embed / lm_head skipped in mask (body-only stitch; matches prior Plan B protocol)

Uploaded: 2026-09-11.