miru-codev3-distill-fuse

Owned distill_fuse static code embedding bag (256-d, Model2Vec) — branch B of miru-codev3-dual.

Distilled from teachers 0.3 · tokenlearn + 0.7 · potion-code-16M-v2 on CornStack texts (code_static.train_distill_fuse). Potion is only a training teacher — not required at inference.

Evaluation

CoIR / MTEB NDCG@10 (×100), dense retrieval only. Self-reported.

Usage

from model2vec import StaticModel

model = StaticModel.from_pretrained("Takara-DS1/miru-codev3-distill-fuse")
emb = model.encode(["def add(a, b): return a + b"])

For the 512-d dual (tokenlearn ⊕ this bag), use Takara-DS1/miru-codev3-dual.

Recreate

See the dual repo reproduce/REPRODUCE.md (step 2: train_distill_fuse --alpha 0.3).

License

MIT