v0.8.0

What's New

• normalize parameter for AttentionLogOddsWeights -- per-signal logit normalization is now a first-class model parameter. When normalize=True, applies per-column min-max normalization in logit space before the weighted sum, equalizing signal scales (same scaling as balanced_log_odds_fusion).

  • __call__: normalizes logit columns across all candidates for a given query
  • fit: accepts optional query_ids parameter to normalize within each query group; without query_ids, normalizes the whole batch as a single group
  • update: normalizes logit columns when input is 2D (assumes same query)

• Broadcasting fix -- AttentionLogOddsWeights.__call__ now correctly broadcasts single query features across batched probability inputs

• Benchmark simplification -- removed the external sigmoid trick from benchmarks/hybrid_beir.py; normalization is now handled internally via normalize=True with per-query query_ids for correct train/inference parity

BEIR Benchmark (NDCG@10)

Method	ArguAna	FiQA	NFCorpus	SciDocs	SciFact	Average
BM25	36.13	25.31	31.82	15.63	68.02	35.38
Convex	40.01	37.10	35.60	19.67	73.37	41.15
Bayesian-Balanced	37.27	40.58	35.73	21.42	72.47	41.50
Attn-NR	37.22	40.53	35.42	21.91	73.24	41.67

Attn-NR achieves the highest zero-shot NDCG@10 average (+6.28 vs BM25).

Install

pip install bayesian-bm25==0.8.0

Full Changelog: https://github.com/cognica-io/bayesian-bm25/blob/main/HISTORY.md#080-2026-03-05