Add `do_query_expansion: false` to load without query expansion
Hello!
Heads up, this PR was AI-generated and human-reviewed.
Pull Request overview
- Add
"do_query_expansion": falsetoconfig_sentence_transformers.json, so the model loads without ColBERT-style[MASK]query expansion.
Details
When this model is loaded through PyLate or Sentence Transformers' MultiVectorEncoder, a ColBERT config with no explicit do_query_expansion flag defaults to query expansion being on. That pads every query out to query_length with [MASK] tokens. If the model was trained without query expansion, those never-trained expansion vectors dominate the MaxSim score for short queries and pull retrieval quality down.
Adding do_query_expansion: false restores the behavior the model was trained with. On the 13 NanoBEIR datasets (mean nDCG@10) this particular model is roughly neutral to the setting (0.598 versus 0.597). We are still adding the flag because the model was trained without query expansion, so this makes the configuration match the training, and it keeps the setting consistent across the SauerkrautLM ColBERT family (the others improve clearly with it).
The change is a single key in config_sentence_transformers.json, nothing else is touched. Happy to adjust the wording or approach if you had something else in mind.
- Tom Aarsen