lecore-bge-assimilated

lecore-bge-assimilated

BAAI/bge-base-en-v1.5 passed through leCore's assimilation + budgeted requantize machinery (Marchenko-Pastur spectral filter, then per-tensor bit-width chosen by mean-embedding-cosine ≥ 0.99 at group-64).

What happened, measured (2026-08-14):

• MP filter: 0 layers filtered (71/73 heavy-tail passthrough) — the predicted no-op on a well-trained encoder.
• Requantize landed at mean 7.62 bits/weight (15×3b, 4×4b, 9×5b, 14×6b, 20×8b, 11 tensors kept fp32 — the encoder resists the ~3-bit the same machinery achieved on a 9B decoder). Packed estimate ~132 MB; this repo ships the dequantized-fp32 container (438 MB) snapped to the quant grid — it loads anywhere bge-base loads.
• Quality: SciFact nDCG@10 0.7388 (full bge: 0.7404); NFCorpus 0.3716 (0.3735); ArguAna 0.6375 (identical); SCIDOCS 0.2152 (0.2172). Zero-lexical-overlap NIAH recall@64 0.8464 — identical to full bge. CPU inference scores bit-identical to GPU.
• Honesty notes: requantize buys size, not CPU matmul speed (inference runs at normal bge-base pace — ~27–31 ms batch-1 query on 8 CPU threads); a packed runtime (e.g. GGUF) would be needed to realize the memory number at inference; uniform 4-bit (~93 MB, 98.4% quality) measured as the better packed-runtime point, uniform 3-bit degrades visibly (0.6922) and should not ship.

Method, full matrix (adversarial decoys, FAISS flat/HNSW, cost-to-first-answer), and every number's provenance: https://benches.openzoo.fun