pair-blend-cache-v3

Private cache of pair-scoring combiner + sample embedding memmaps for internal reuse. Base encoders are upstream ST/BGE/E5/GTE.

This is not a fine-tuned Sentence-Transformer checkpoint. Upstream base model weights remain on their original Hugging Face repos. This repo holds:

1. Combiner weights — scikit-learn LogisticRegression mixing fuzzy + semantic cosines (weights/logistic_combiner.joblib + weights/combiner_meta.json).
2. Cached embeddings (fp16 .npy + _ids.json) for a local sample holdout — S1 (15k) and src (104k) for each of five base encoders.
3. Run metrics — metrics.json, f05_by_slice.json, RESULTS.md from the holdout that reached macro F₀.₅ ≈ 0.925.

Base encoders (upstream HF ids)

Key	Hugging Face id	Notes
mpnet	sentence-transformers/all-mpnet-base-v2
bge	BAAI/bge-base-en-v1.5
e5	intfloat/e5-base-v2	both sides prefixed query:
gte	thenlper/gte-base
para_mpnet	sentence-transformers/paraphrase-mpnet-base-v2

Pipeline (sample holdout)

1. Fuzzy retrieve: char 3-gram TF-IDF (name + address, addr weight 0.85), country-blocked, TOP_K=20, drop fuzzy_score < 0.3.
2. Ensemble cosines: per-model semantic similarity → sem_mpnet, sem_bge, sem_e5, sem_gte, sem_para_mpnet.
3. Combiner: LogisticRegression(max_iter=500, class_weight="balanced", solver="lbfgs", random_state=42, C=1.0) on
   ['fuzzy_score', 'sem_mpnet', 'sem_bge', 'sem_e5', 'sem_gte', 'sem_para_mpnet', 'same_country', 'name_len_ratio'].
4. Decision: each src → at most one S1 (argmax combo_score ≥ 0.85); invert to S1→src sets.

Split protocol: SEED=42, holdout 20%, matched-train permutation seed 42+7.

Holdout metrics

Metric	Value
Macro F₀.₅	0.9254
Macro P	0.9263
Macro R	0.9341

Combiner coefficients (refit-verified)

Feature	Coef
fuzzy_score	+7.82
sem_mpnet	−8.47
sem_bge	+26.66
sem_e5	+43.24
sem_gte	+0.76
sem_para_mpnet	−3.69
same_country	−24.00
name_len_ratio	+0.79
intercept	−30.73

Layout

weights/
  logistic_combiner.joblib
  combiner_meta.json
embeddings/
  s1_<model>.npy / _ids.json
  src_<model>.npy / _ids.json
metrics.json
f05_by_slice.json
RESULTS.md