BSCA BGE-Micro-v2 Gold v3 Multidomain
Embedding model for stripped pseudo-C → source retrieval.
- Model ID:
Labradorlabs/bsca-bge-micro-v2-contrastive-v15-gold-v3-multidomain-384
- Backbone:
TaylorAI/bge-micro-v2
- Initialization / training mode: weights-restart from
artifacts/reference_v2/models/gold-scale-v2-seed44; optimizer, scheduler, and RNG state were reinitialized (fresh-or-weights-restart)
- Dimension / maximum tokens:
384 / 384
- Development-only selected seed:
44 (R@1 0.3186, MRR 0.3553)
- Frozen test result: R@1 0.3294, R@10 0.6488, MRR 0.4431
- Training data:
Labradorlabs/bsca-binary-source-gold-v3-multidomain
Embed query pseudo-C and source code with attention-mask mean pooling followed
by L2 normalization. Use the exact model revision recorded by the associated
ReferenceDB vector manifest; do not combine it with vectors from another
model revision.
Scope is limited to the accompanying Gold distribution: ELF/Linux and PE
x86_64 Windows-GNU direct-TU views. The model is not promoted by an external
blind SCA gate yet.
