Sourced from sentence-transformers's releases.
v5.6.0 - Fixes for Causal LM Rerankers, Hard-Negative Mining, and More
This minor version is a correctness- and robustness-focused release. It fixes a silent scoring bug for causal-LM rerankers, corrects several hard-negative mining and GIST loss edge cases, restores TSDAE on
transformersv5, and adds Apple Silicon (MPS) support for the cached losses.The headline fix affects chat-template models that read the final token position, i.e. causal-LM rerankers (like
Qwen3-Reranker) and last-token-pooling embedders: when an over-long input was truncated, the chat template's trailing suffix (e.g. the assistant prefill the model scores from) was silently dropped, producing wrong scores with no error. There's also a forward-looking deprecation: loading local custom code withouttrust_remote_code=Truenow warns, and will require it from v6.0.Install this version with
# Training + Inference pip install sentence-transformers[train]==5.6.0Inference only, use one of:
pip install sentence-transformers==5.6.0 pip install sentence-transformers[onnx-gpu]==5.6.0 pip install sentence-transformers[onnx]==5.6.0 pip install sentence-transformers[openvino]==5.6.0
Multimodal dependencies (optional):
pip install sentence-transformers[image]==5.6.0 pip install sentence-transformers[audio]==5.6.0 pip install sentence-transformers[video]==5.6.0
Or combine as needed:
pip install sentence-transformers[train,onnx,image]==5.6.0
Fixed silently wrong scores when truncation drops chat-template suffixes (#3787)
Chat-template models render the full conversation to a flat string before tokenizing, so when the rendered input is longer than the tokenizer's
model_max_length, the tokenizer truncates it from the right and drops the template's trailing suffix: the fixed tokens a template appends after the content, e.g. a prompt, instruction,[/INST], or a trailing EOS. For models that read the final token position, this silently corrupted the result:
- causal-LM rerankers (e.g.
Qwen/Qwen3-Reranker-0.6B) score a pair from the last token'syes/nologits, and- last-token-pooling embedders read the final hidden state.
When the suffix was truncated away, that final position landed mid-document instead of after the prefill, so the score or embedding came from the wrong place.
Transformer.preprocessnow detects when truncation drops the suffix and splices it back onto the tail of each truncated row. Because the fix lives in the shared baseTransformer, it applies acrossSentenceTransformer,CrossEncoder, andSparseEncoder. It's enabled by default and saved to the model configuration. Passprocessing_kwargs={"chat_template": {"restore_suffix": False}}to opt back into raw truncation.Hard-negative mining and GIST loss correctness (#3821, #3817, #3816)
A trio of correctness and scalability fixes for hard-negative mining and the GIST losses:
- Sign-independent relative margin:
mine_hard_negatives(relative_margin=...)and themargin_strategy="relative"branch ofGISTEmbedLoss/CachedGISTEmbedLossused a multiplicative threshold (positive * (1 - margin)) that only behaves correctly when the positive-pair similarity is positive. When that similarity was negative, the threshold moved the wrong way and let through false negatives: candidates more similar to the anchor than the true positive. The threshold is nowpositive - |positive| * margin, identical to before for positive scores but correct for negative ones.- Distributed positive masking in the GIST losses: with
gather_across_devices=Trueand a non-zeromargin, the false-negative suppression mask protected the wrong columns on ranks beyond the first (it ignored the per-rank offset into the gathered batch), which set the true positive's logit to-infand produced a+infloss. The mask now accounts for the cross-rank offset, so multi-GPU GIST training stays finite.- Memory-bounded mining without FAISS:
mine_hard_negatives(use_faiss=False)(the default) materialized the full(queries × corpus)similarity matrix at once, which could OOM on large corpora. It now batches over the query axis (controlled byfaiss_batch_size, default 16384), bounding peak memory while producing identical results.TSDAE weight tying restored on
transformersv5 (#3781)
transformersv5 removed the privatePreTrainedModel._tie_encoder_decoder_weightshelper thatDenoisingAutoEncoderLoss(TSDAE) used to tie its separate encoder and decoder. As a stopgap, v5.5 raised aRuntimeErrorfor the defaulttie_encoder_decoder=Trueontransformers >= 5.0.0, effectively breaking TSDAE there unless you pinned an oldertransformersor disabled tying. TSDAE now ships its own tying routine that shares storage between encoder and decoder, so it works on bothtransformers<5 and >=5 with the default settings.Deprecation: loading local custom code without
trust_remote_code(#3807)
... (truncated)
9c73df3
Release v5.6.0222b052
[fix] Don't override device_map placement with the device argument (#3823)a38a6bf
Fix causal LM reranker scoring when max_length truncates chat-template
suffix...1812103
[fix] Make relative margin sign-independent in mining and GIST losses
(#3821)ae1acc3
Warn when loading local custom code without trust_remote_code (#3807)429cf5d
[fix] Support MPS in the cached losses' RandContext (#3812)77fdbff
[fix] fix MPS errors (#3818)bfba988
[fix] Fix positive masking in GIST losses with multi-GPU +
gather_across_devi...29e382b
[docs] Fix doc build problems (part 1) (#3811)d16e6bf
[fix] Avoid materializing the full similarity matrix in
mine_hard_negatives w...