Three of my fixes landed in sentence-transformers v5.6.0, all in the hard-negative mining and GIST loss code paths. Two fix wrong training behavior in the mining and GIST losses; one removes an out-of-memory ceiling.
#3816: mining materialized the full similarity matrix
mine_hard_negatives without FAISS computed the full query-by-corpus similarity matrix in one allocation, and the memory cost scales with the product of the two sizes. The fix mirrors the FAISS branch: iterate over query chunks, keeping only the top candidates per chunk. In my local CPU benchmark (all-MiniLM-L6-v2, synthetic 10k queries by 100k documents), peak RSS dropped from 4.56 GB to 1.09 GB, and the mined output was identical in equivalence checks. Mined results are unchanged by design; this fix removes an out-of-memory ceiling, not a bias. PR #3816
#3817: multi-GPU GIST masked the wrong positives
GIST losses use a guide model to mask false negatives. With gather_across_devices=True, embeddings from all ranks are concatenated, so each anchor’s true positive sits at an offset determined by the process rank. The positive mask did not apply that offset: on every rank above zero the mask pointed at another rank’s rows. With a nonzero margin the unprotected true-positive logit became -inf, which yields an infinite loss term; with the default margin the wrong rows were filtered. Single-GPU runs were unaffected; the failure exists only with gather_across_devices=True. The fix applies the rank offset in both GISTEmbedLoss and CachedGISTEmbedLoss. The merged regression tests simulate rank 1 of a 2-process world in a single process; I also verified the red/green behavior locally on a 2-process gloo CPU run. PR #3817
#3821: the relative margin flipped for negative scores
Mining and GIST losses discarded negatives scoring above positive_score * (1 - margin). For positive scores this is a threshold slightly below the positive. For negative positive-pair scores it lands above the positive: with a positive score of -0.50 and margin=0.05, the threshold is -0.475, so a negative at -0.49, closer to the anchor than the positive itself, survives the filter. The fix is a sign-independent form, positive_score - abs(positive_score) * margin, identical for positive scores and correct for negative ones, applied in the three places that shared the expression (mine_hard_negatives, GISTEmbedLoss, CachedGISTEmbedLoss). PR #3821
Regression properties
Each fix is pinned by a property test: mined output identity across memory strategies (#3816), single-GPU equivalence for the distributed mask (#3817), and threshold ordering across score signs (#3821). All three fail on the previous code and pass after the fixes.