Qwen

/

Qwen3-VL-Reranker-8B

Like 179

Follow

Qwen 111k

Loading the model with Sentence Transformers currently fails:

from sentence_transformers import CrossEncoder

model = CrossEncoder("Qwen/Qwen3-VL-Reranker-8B")
# TypeError: LogitScore.__init__() missing 1 required positional argument: 'true_token_id'

modules.json points the score head to 1_CausalScoreHead, but the folder in the repository is 1_LogitScore
(containing {"true_token_id": 9693, "false_token_id": 2152}). Since no config is found for the head, LogitScore is
constructed without true_token_id. Qwen/Qwen3-VL-Reranker-2B uses "path": "1_LogitScore" and loads fine.

This PR points the score head to the existing folder:

   {
     "idx": 1,
     "name": "1",
-    "path": "1_CausalScoreHead",
+    "path": "1_LogitScore",
     "type": "sentence_transformers.cross_encoder.modules.logit_score.LogitScore"
   }

With this change the model loads and the head gets true_token_id=9693 / false_token_id=2152.

Reproduced on b212dc8c91a8164aef1ea2de9c1a867611e75c04 (current main) with sentence-transformers 6.1.0 / transformers 5.19.0,
and with sentence-transformers 5.7.0.