Fix bidirlm omni double frame sampling (#4700)

* fix: disable inner video sampler in BidirLM-Omni encoder

VideoCollator already samples frames at the requested fps, then drops
the video metadata before handing a 4D tensor to the model. With
do_sample_frames=True (the BidirLM-Omni video_preprocessor_config
default), Qwen3VLVideoProcessor.sample_frames hits the metadata.fps=24
fallback and downsamples to ~T/12, throwing away most of the frames
the outer collator extracted. Set do_sample_frames=False on the loaded
processor at construction so sampling happens exactly once.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* update method for setting frames sampling

* simplify

* move imports inside

---------

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-authored-by: Roman Solomatin <36135455+Samoed@users.noreply.github.com>

Author: 3 people

mteb/models/model_implementations/bidirlm_omni_models.py

@@ -94,6 +94,23 @@ def __init__(
         max_samples: int | None = None,
         **kwargs: Any,
     ) -> None:
+        from transformers import AutoVideoProcessor
+
+        processor_kwargs = kwargs.get("processor_kwargs", {})
+
+        # VideoCollator already samples frames; skip inner sampling.
+        video_processor = AutoVideoProcessor.from_pretrained(
+            model_name,
+            revision=revision,
+            trust_remote_code=True,
+            do_sample_frames=False,
+            **kwargs,
+        )
+        processor_kwargs |= {
+            "video_processor": video_processor,
+        }
+
+        kwargs["processor_kwargs"] = processor_kwargs
         self.model = SentenceTransformer(
             model_name,
             revision=revision,