fix logit bias

README.md

@@ -178,7 +178,7 @@ Compared to `jina-reranker-v2-base-multilingual`, `jina-reranker-m0` significant
     image_pairs = [[query, doc] for doc in documents]
 
     scores = model.compute_score(image_pairs, max_length=2048, doc_type="image")
-    # [0.8576154708862305, 0.9356858730316162, 0.8496521711349487, 0.8664582967758179]
+    # [0.766852855682373, 0.9265167713165283, 0.7554926872253418, 0.7858350276947021]
     ```
 
     **B. Textual Documents Reranking**
@@ -201,7 +201,7 @@ Compared to `jina-reranker-v2-base-multilingual`, `jina-reranker-m0` significant
     The scores will be a list of floats, where each float represents the relevance score of the corresponding document to the query. Higher scores indicate higher relevance.
     For instance the returning scores in this case will be:
     ```bash
-    [0.9127850532531738, 0.8384682536125183, 0.8870794177055359, 0.842738926410675]
+    [0.8778123259544373, 0.7254930734634399, 0.8271589875221252, 0.7437640428543091]
     ```
 
     **C. Image Querying for Textual Documents**
@@ -221,7 +221,7 @@ Compared to `jina-reranker-v2-base-multilingual`, `jina-reranker-m0` significant
     image_pairs = [[doc, query] for doc in documents]
     scores = model.compute_score(image_pairs, max_length=2048, doc_type="text")
 
-    # [0.9048659801483154, 0.8266222476959229, 0.8326289653778076, 0.9075747132301331]
+    # [0.8673955798149109, 0.6999112367630005, 0.7031826972961426, 0.8744207620620728]
     ```
 
 # Model Performance

modeling.py

@@ -10,7 +10,7 @@ from transformers.image_utils import load_image
 
 logger = logging.getLogger(__name__)
 
-LOGIT_SCALE = 0.68
+LOGIT_BIAS = 1.45  # logit bias for sigmoid normalization
 
 def load_images(images, lazy_load: bool = True):
     # Disable PIL DecompositionBomb threshold for reading large images.
@@ -123,7 +123,7 @@ class JinaVLForRanking(Qwen2VLForConditionalGeneration):
         pairs: Union[List[Tuple[str, str]], Tuple[str, str]],
         batch_size: int = 8,
         max_length: int = 10240,
-        max_query_length: int = 1024,
+        max_query_length: int = 512,
         max_doc_length: Optional[int] = None,
         query_type: str = 'text',
         doc_type: str = 'text',
@@ -219,7 +219,7 @@ class JinaVLForRanking(Qwen2VLForConditionalGeneration):
             scores = self.forward(**batch).view(-1).cpu().float().numpy()
 
             # normalize scores to [0, 1] with sigmoid with a scale
-            scores = 1.0 / (1.0 + np.exp(-scores * LOGIT_SCALE))
+            scores = 1.0 / (1.0 + np.exp(-(scores - LOGIT_BIAS)))
 
             all_scores.extend(scores.tolist())