Expand description
Supervised fitting for query-level Bayesian BM25 calibration.
The fitted posterior is the same transform used by
BayesianBM25Scorer: sigmoid(alpha * (raw_bm25_score - beta)).
Any intercept the labels call for is absorbed into beta. The
base_rate is not a model term: it tracks the observed positive
label rate as an exponential moving average, estimating the corpus
relevance prior that fusion applies exactly once.