Fix DetailsLogger hashing with xxhash 4.0 (#1402)
* Encode strings before hashing with xxhash
xxhash 4.0 no longer accepts str input to xxh64 (TypeError: Strings must be
encoded before hashing), which breaks DetailsLogger and turns tests red on
main. Encode to UTF-8 explicitly, matching xxhash 3.x's implicit behaviour so
the resulting hashes are unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Update Qwen2.5-VL-3B mmmu_pro reference for sample 16
With current dependencies the model answers D instead of B on sample 16
(gold is C), so extractive_match stays 0 and aggregate scores are unchanged.
The drift was hidden until now because the slow tests crashed on xxhash.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Nathan Habib <nathan_habib@Nathans-MacBook-Pro.local>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
