mteb 2.5.2__py3-none-any.whl → 2.5.4__py3-none-any.whl

mteb/models/cache_wrappers/cache_backends/numpy_cache.py

@@ -1,11 +1,11 @@
 import json
 import logging
+import warnings
 from pathlib import Path
+from typing import Any
 
 import numpy as np
 
-from mteb.types import BatchedInput
-
 from ._hash_utils import _hash_item
 
 logger = logging.getLogger(__name__)
@@ -14,7 +14,7 @@ logger = logging.getLogger(__name__)
 class NumpyCache:
     """Generic vector cache for both text and images."""
 
-    def __init__(self, directory: str | Path, initial_vectors: int = 100000):
+    def __init__(self, directory: str | Path, initial_vectors: int = 100_000):
         self.directory = Path(directory)
         self.directory.mkdir(parents=True, exist_ok=True)
         self.vectors_file = self.directory / "vectors.npy"
@@ -27,7 +27,7 @@ class NumpyCache:
         logger.info(f"Initialized VectorCacheMap in directory: {self.directory}")
         self._initialize_vectors_file()
 
-    def add(self, item: list[BatchedInput], vectors: np.ndarray) -> None:
+    def add(self, items: list[dict[str, Any]], vectors: np.ndarray) -> None:
         """Add a vector to the cache."""
         try:
             if self.vector_dim is None:
@@ -38,12 +38,17 @@ class NumpyCache:
                 self._save_dimension()
                 logger.info(f"Initialized vector dimension to {self.vector_dim}")
 
-            for item, vec in zip(item, vectors):
+            if self.vectors is None:
+                raise RuntimeError(
+                    "Vectors file not initialized. Call _initialize_vectors_file() first."
+                )
+
+            for item, vec in zip(items, vectors):
                 item_hash = _hash_item(item)
                 if item_hash in self.hash_to_index:
-                    logger.warning(
-                        "Hash collision or duplicate item. Overwriting existing vector."
-                    )
+                    msg = f"Hash collision or duplicate item for hash {item_hash}. Overwriting existing vector."
+                    logger.warning(msg)
+                    warnings.warn(msg)
                     index = self.hash_to_index[item_hash]
                 else:
                     index = len(self.hash_to_index)
@@ -74,18 +79,26 @@ class NumpyCache:
                 shape=(self.initial_vectors, self.vector_dim),
             )
         else:
-            self.vectors = np.memmap(self.vectors_file, dtype="float32", mode="r+")
-            self.vectors = self.vectors.reshape(-1, self.vector_dim)
+            self.vectors = np.memmap(
+                self.vectors_file,
+                dtype="float32",
+                mode="r+",
+                shape=(-1, self.vector_dim),
+            )
         logger.info(f"Vectors file initialized with shape: {self.vectors.shape}")
 
     def _double_vectors_file(self) -> None:
+        if self.vectors is None or self.vector_dim is None:
+            raise RuntimeError(
+                "Vectors file not initialized. Call _initialize_vectors_file() first."
+            )
         current_size = len(self.vectors)
         new_size = current_size * 2
         logger.info(f"Doubling vectors file from {current_size} to {new_size} vectors")
         self.vectors.flush()
         new_vectors = np.memmap(
-            self.vectors_file,
-            dtype="float32",
+            str(self.vectors_file),
+            dtype=np.float32,
             mode="r+",
             shape=(new_size, self.vector_dim),
         )
@@ -107,9 +120,9 @@ class NumpyCache:
                 f"Loaded vector dimension {self.vector_dim} from {self.dimension_file}"
             )
         else:
-            logger.warning(
-                "Dimension file not found. Vector dimension remains uninitialized."
-            )
+            msg = "Dimension file not found. Vector dimension remains uninitialized."
+            logger.warning(msg)
+            warnings.warn(msg)
 
     def save(self) -> None:
         """Persist VectorCacheMap to disk."""
@@ -146,25 +159,30 @@ class NumpyCache:
 
                 if self.vector_dim is not None:
                     self.vectors = np.memmap(
-                        self.vectors_file, dtype="float32", mode="r+"
+                        self.vectors_file,
+                        dtype="float32",
+                        mode="r+",
+                        shape=(-1, self.vector_dim),
                     )
-                    self.vectors = self.vectors.reshape(-1, self.vector_dim)
                     logger.info(f"Loaded vectors file with shape: {self.vectors.shape}")
                 else:
-                    logger.warning(
-                        "Vector dimension not set. Unable to load vectors file."
-                    )
+                    msg = "Vector dimension not set. Unable to load vectors file."
+                    logger.warning(msg)
+                    warnings.warn(msg)
                 logger.info(f"Loaded VectorCacheMap from {self.directory}")
             else:
-                logger.warning(
-                    "No existing files found. Initialized empty VectorCacheMap."
-                )
+                msg = "No existing files found. Initialized empty VectorCacheMap."
+                logger.warning(msg)
+                warnings.warn(msg)
         except Exception as e:
             logger.error(f"Error loading VectorCacheMap: {str(e)}")
             raise
 
-    def get_vector(self, item: BatchedInput) -> np.ndarray | None:
+    def get_vector(self, item: dict[str, Any]) -> np.ndarray | None:
         """Retrieve vector from index by hash."""
+        if self.vectors is None:
+            return None
+
         try:
             item_hash = _hash_item(item)
             if item_hash not in self.hash_to_index:
@@ -176,7 +194,7 @@ class NumpyCache:
             logger.error(f"Error retrieving vector for item: {str(e)}")
             raise
 
-    def __contains__(self, item: BatchedInput) -> bool:
+    def __contains__(self, item: dict[str, Any]) -> bool:
         return _hash_item(item) in self.hash_to_index
 
     def __del__(self):

mteb/models/cache_wrappers/cache_wrapper.py

@@ -90,9 +90,9 @@ class CachedEmbeddingWrapper:
         try:
             cache = self._get_or_create_cache(task_name)
 
-            uncached_items: list[BatchedInput] = []
+            uncached_items: list[dict[str, Any]] = []
             uncached_indices: list[int] = []
-            all_items = inputs.dataset
+            all_items: Dataset = inputs.dataset
             cached_vectors: dict[int, np.ndarray] = {}
 
             for i, item in enumerate(all_items):

mteb/models/get_model_meta.py

@@ -93,7 +93,14 @@ def get_model(
     meta = get_model_meta(model_name, revision)
     model = meta.load_model(**kwargs)
 
-    model.mteb_model_meta = meta  # type: ignore
+    if kwargs:
+        logger.info(
+            f"Model '{model_name}' loaded with additional arguments: {list(kwargs.keys())}"
+        )
+        meta = meta.model_copy(deep=True)
+        meta.loader_kwargs |= kwargs
+
+    model.mteb_model_meta = meta  # type: ignore[misc]
     return model
 
 

mteb/models/instruct_wrapper.py

@@ -17,7 +17,7 @@ logger = logging.getLogger(__name__)
 def instruct_wrapper(
     model_name_or_path: str,
     mode: str,
-    instruction_template: str | Callable[[str], str] | None = None,
+    instruction_template: str | Callable[[str, PromptType | None], str] | None = None,
     **kwargs,
 ):
     """Instruct wrapper for models. Uses GritLM to pass instructions to the model.
@@ -40,7 +40,9 @@ def instruct_wrapper(
             self,
             model_name_or_path: str,
             mode: str,
-            instruction_template: str | Callable[[str, PromptType], str] | None = None,
+            instruction_template: str
+            | Callable[[str, PromptType | None], str]
+            | None = None,
             **kwargs,
         ):
             if (
@@ -82,8 +84,11 @@ def instruct_wrapper(
             logger.info(
                 f"Using instruction: '{instruction}' for task: '{task_metadata.name}'"
             )
-            embeddings = super().encode(
-                _inputs, instruction=instruction, *args, **kwargs
+            embeddings = super().encode(  # type: ignore[safe-super]
+                _inputs,  # type: ignore[arg-type]
+                instruction=instruction,
+                *args,
+                **kwargs,
             )
             if isinstance(embeddings, torch.Tensor):
                 # sometimes in kwargs can be return_tensors=True
@@ -141,7 +146,7 @@ class InstructSentenceTransformerModel(AbsEncoder):
             )
 
         self.instruction_template = instruction_template
-        tokenizer_params = {}
+        tokenizer_params: dict[str, Any] = {}
         if add_eos_token:
             tokenizer_params["add_eos_token"] = add_eos_token
         if max_seq_length is not None:
@@ -193,6 +198,7 @@ class InstructSentenceTransformerModel(AbsEncoder):
             The encoded input in a numpy array or torch tensor of the shape (Number of sentences) x (Embedding dimension).
         """
         sentences = [text for batch in inputs for text in batch["text"]]
+        instruction: str | None
         instruction = self.get_task_instruction(task_metadata, prompt_type)
 
         # to passage prompts won't be applied to passages

mteb/models/model_implementations/andersborges.py

@@ -4,7 +4,7 @@ from mteb.models.model_implementations.model2vec_models import Model2VecModel
 from mteb.models.model_meta import ModelMeta, ScoringFunction
 
 model2vecdk = ModelMeta(
-    loader=Model2VecModel,  # type: ignore
+    loader=Model2VecModel,
     name="andersborges/model2vecdk",
     model_type=["dense"],
     languages=["dan-Latn"],
@@ -35,7 +35,7 @@ model2vecdk = ModelMeta(
 
 
 model2vecdk_stem = ModelMeta(
-    loader=Model2VecModel,  # type: ignore
+    loader=Model2VecModel,
     name="andersborges/model2vecdk-stem",
     model_type=["dense"],
     languages=["dan-Latn"],

mteb/models/model_implementations/blip_models.py

@@ -128,7 +128,7 @@ class BLIPModel(AbsEncoder):
 
 # in descending order of usage (downloads from huggingface)
 blip_image_captioning_large = ModelMeta(
-    loader=BLIPModel,  # type: ignore
+    loader=BLIPModel,
     name="Salesforce/blip-image-captioning-large",
     model_type=["dense"],
     languages=["eng-Latn"],
@@ -156,7 +156,7 @@ blip_image_captioning_large = ModelMeta(
 )
 
 blip_image_captioning_base = ModelMeta(
-    loader=BLIPModel,  # type: ignore
+    loader=BLIPModel,
     name="Salesforce/blip-image-captioning-base",
     model_type=["dense"],
     languages=["eng-Latn"],
@@ -185,7 +185,7 @@ blip_image_captioning_base = ModelMeta(
 
 
 blip_vqa_base = ModelMeta(
-    loader=BLIPModel,  # type: ignore
+    loader=BLIPModel,
     name="Salesforce/blip-vqa-base",
     model_type=["dense"],
     languages=["eng-Latn"],
@@ -212,7 +212,7 @@ blip_vqa_base = ModelMeta(
 )
 
 blip_vqa_capfilt_large = ModelMeta(
-    loader=BLIPModel,  # type: ignore
+    loader=BLIPModel,
     name="Salesforce/blip-vqa-capfilt-large",
     model_type=["dense"],
     languages=["eng-Latn"],
@@ -239,7 +239,7 @@ blip_vqa_capfilt_large = ModelMeta(
 )
 
 blip_itm_base_coco = ModelMeta(
-    loader=BLIPModel,  # type: ignore
+    loader=BLIPModel,
     name="Salesforce/blip-itm-base-coco",
     model_type=["dense"],
     languages=["eng-Latn"],
@@ -266,7 +266,7 @@ blip_itm_base_coco = ModelMeta(
 )
 
 blip_itm_large_coco = ModelMeta(
-    loader=BLIPModel,  # type: ignore
+    loader=BLIPModel,
     name="Salesforce/blip-itm-large-coco",
     model_type=["dense"],
     languages=["eng-Latn"],
@@ -294,7 +294,7 @@ blip_itm_large_coco = ModelMeta(
 )
 
 blip_itm_base_flickr = ModelMeta(
-    loader=BLIPModel,  # type: ignore
+    loader=BLIPModel,
     name="Salesforce/blip-itm-base-flickr",
     model_type=["dense"],
     languages=["eng-Latn"],
@@ -322,7 +322,7 @@ blip_itm_base_flickr = ModelMeta(
 )
 
 blip_itm_large_flickr = ModelMeta(
-    loader=BLIPModel,  # type: ignore
+    loader=BLIPModel,
     name="Salesforce/blip-itm-large-flickr",
     model_type=["dense"],
     languages=["eng-Latn"],

mteb/models/model_implementations/bm25.py

@@ -113,7 +113,7 @@ def bm25_loader(model_name, **kwargs) -> SearchProtocol:
 
         def encode(self, texts: list[str]):
             """Encode input text as term vectors"""
-            return bm25s.tokenize(texts, stopwords=self.stopwords, stemmer=self.stemmer)  # type: ignore
+            return bm25s.tokenize(texts, stopwords=self.stopwords, stemmer=self.stemmer)
 
     return BM25Search(**kwargs)
 

mteb/models/model_implementations/clip_models.py

@@ -115,7 +115,7 @@ CLIP_CITATION = """
 
 
 clip_vit_large_patch14 = ModelMeta(
-    loader=CLIPModel,  # type: ignore
+    loader=CLIPModel,
     name="openai/clip-vit-large-patch14",
     model_type=["dense"],
     languages=["eng-Latn"],
@@ -139,7 +139,7 @@ clip_vit_large_patch14 = ModelMeta(
 )
 
 clip_vit_base_patch32 = ModelMeta(
-    loader=CLIPModel,  # type: ignore
+    loader=CLIPModel,
     name="openai/clip-vit-base-patch32",
     model_type=["dense"],
     languages=["eng-Latn"],
@@ -163,7 +163,7 @@ clip_vit_base_patch32 = ModelMeta(
 )
 
 clip_vit_base_patch16 = ModelMeta(
-    loader=CLIPModel,  # type: ignore
+    loader=CLIPModel,
     name="openai/clip-vit-base-patch16",
     model_type=["dense"],
     languages=["eng-Latn"],

mteb/models/model_implementations/cohere_models.py

@@ -222,7 +222,7 @@ class CohereTextEmbeddingModel(AbsEncoder):
     ) -> None:
         requires_package(self, "cohere", model_name, "pip install 'mteb[cohere]'")
 
-        import cohere  # type: ignore
+        import cohere
 
         self.model_name = model_name.removeprefix("Cohere/Cohere-")
         self.sep = sep

mteb/models/model_implementations/cohere_v.py

@@ -378,7 +378,7 @@ def cohere_v_loader(model_name, **kwargs):
 
 
 cohere_mult_3 = ModelMeta(
-    loader=cohere_v_loader,  # type: ignore
+    loader=cohere_v_loader,
     loader_kwargs={"model_name": "embed-multilingual-v3.0"},
     name="cohere/embed-multilingual-v3.0",
     model_type=["dense"],
@@ -402,7 +402,7 @@ cohere_mult_3 = ModelMeta(
 )
 
 cohere_eng_3 = ModelMeta(
-    loader=cohere_v_loader,  # type: ignore
+    loader=cohere_v_loader,
     loader_kwargs={"model_name": "embed-english-v3.0"},
     name="cohere/embed-english-v3.0",
     model_type=["dense"],

mteb/models/model_implementations/dino_models.py

@@ -104,7 +104,7 @@ dinov2_training_datasets = set(
 
 
 dinov2_small = ModelMeta(
-    loader=DINOModel,  # type: ignore
+    loader=DINOModel,
     name="facebook/dinov2-small",
     model_type=["dense"],
     languages=["eng-Latn"],
@@ -125,7 +125,7 @@ dinov2_small = ModelMeta(
     use_instructions=False,
     training_datasets=dinov2_training_datasets,
     citation="""@misc{oquab2023dinov2,
-      title={DINOv2: Learning Robust Visual Features without Supervision}, 
+      title={DINOv2: Learning Robust Visual Features without Supervision},
       author={Maxime Oquab and Timothée Darcet and Théo Moutakanni and Huy Vo and Marc Szafraniec and Vasil Khalidov and Pierre Fernandez and Daniel Haziza and Francisco Massa and Alaaeldin El-Nouby and Mahmoud Assran and Nicolas Ballas and Wojciech Galuba and Russell Howes and Po-Yao Huang and Shang-Wen Li and Ishan Misra and Michael Rabbat and Vasu Sharma and Gabriel Synnaeve and Hu Xu and Hervé Jegou and Julien Mairal and Patrick Labatut and Armand Joulin and Piotr Bojanowski},
       year={2023},
       eprint={2304.07193},
@@ -135,7 +135,7 @@ dinov2_small = ModelMeta(
 )
 
 dinov2_base = ModelMeta(
-    loader=DINOModel,  # type: ignore
+    loader=DINOModel,
     name="facebook/dinov2-base",
     model_type=["dense"],
     languages=["eng-Latn"],
@@ -156,7 +156,7 @@ dinov2_base = ModelMeta(
     use_instructions=False,
     training_datasets=dinov2_training_datasets,
     citation="""@misc{oquab2023dinov2,
-      title={DINOv2: Learning Robust Visual Features without Supervision}, 
+      title={DINOv2: Learning Robust Visual Features without Supervision},
       author={Maxime Oquab and Timothée Darcet and Théo Moutakanni and Huy Vo and Marc Szafraniec and Vasil Khalidov and Pierre Fernandez and Daniel Haziza and Francisco Massa and Alaaeldin El-Nouby and Mahmoud Assran and Nicolas Ballas and Wojciech Galuba and Russell Howes and Po-Yao Huang and Shang-Wen Li and Ishan Misra and Michael Rabbat and Vasu Sharma and Gabriel Synnaeve and Hu Xu and Hervé Jegou and Julien Mairal and Patrick Labatut and Armand Joulin and Piotr Bojanowski},
       year={2023},
       eprint={2304.07193},
@@ -166,7 +166,7 @@ dinov2_base = ModelMeta(
 )
 
 dinov2_large = ModelMeta(
-    loader=DINOModel,  # type: ignore
+    loader=DINOModel,
     name="facebook/dinov2-large",
     model_type=["dense"],
     languages=["eng-Latn"],
@@ -187,7 +187,7 @@ dinov2_large = ModelMeta(
     use_instructions=False,
     training_datasets=dinov2_training_datasets,
     citation="""@misc{oquab2023dinov2,
-      title={DINOv2: Learning Robust Visual Features without Supervision}, 
+      title={DINOv2: Learning Robust Visual Features without Supervision},
       author={Maxime Oquab and Timothée Darcet and Théo Moutakanni and Huy Vo and Marc Szafraniec and Vasil Khalidov and Pierre Fernandez and Daniel Haziza and Francisco Massa and Alaaeldin El-Nouby and Mahmoud Assran and Nicolas Ballas and Wojciech Galuba and Russell Howes and Po-Yao Huang and Shang-Wen Li and Ishan Misra and Michael Rabbat and Vasu Sharma and Gabriel Synnaeve and Hu Xu and Hervé Jegou and Julien Mairal and Patrick Labatut and Armand Joulin and Piotr Bojanowski},
       year={2023},
       eprint={2304.07193},
@@ -197,7 +197,7 @@ dinov2_large = ModelMeta(
 )
 
 dinov2_giant = ModelMeta(
-    loader=DINOModel,  # type: ignore
+    loader=DINOModel,
     name="facebook/dinov2-giant",
     model_type=["dense"],
     languages=["eng-Latn"],
@@ -218,7 +218,7 @@ dinov2_giant = ModelMeta(
     use_instructions=False,
     training_datasets=dinov2_training_datasets,
     citation="""@misc{oquab2023dinov2,
-      title={DINOv2: Learning Robust Visual Features without Supervision}, 
+      title={DINOv2: Learning Robust Visual Features without Supervision},
       author={Maxime Oquab and Timothée Darcet and Théo Moutakanni and Huy Vo and Marc Szafraniec and Vasil Khalidov and Pierre Fernandez and Daniel Haziza and Francisco Massa and Alaaeldin El-Nouby and Mahmoud Assran and Nicolas Ballas and Wojciech Galuba and Russell Howes and Po-Yao Huang and Shang-Wen Li and Ishan Misra and Michael Rabbat and Vasu Sharma and Gabriel Synnaeve and Hu Xu and Hervé Jegou and Julien Mairal and Patrick Labatut and Armand Joulin and Piotr Bojanowski},
       year={2023},
       eprint={2304.07193},
@@ -253,7 +253,7 @@ webssl_dino300m_full2b = ModelMeta(
     use_instructions=False,
     training_datasets=webssl_dino_training_datasets,
     citation="""@article{fan2025scaling,
-  title={Scaling Language-Free Visual Representation Learning}, 
+  title={Scaling Language-Free Visual Representation Learning},
   author={David Fan and Shengbang Tong and Jiachen Zhu and Koustuv Sinha and Zhuang Liu and Xinlei Chen and Michael Rabbat and Nicolas Ballas and Yann LeCun and Amir Bar and Saining Xie},
   year={2025},
   eprint={2504.01017},
@@ -284,7 +284,7 @@ webssl_dino1b_full2b = ModelMeta(
     use_instructions=False,
     training_datasets=webssl_dino_training_datasets,
     citation="""@article{fan2025scaling,
-  title={Scaling Language-Free Visual Representation Learning}, 
+  title={Scaling Language-Free Visual Representation Learning},
   author={David Fan and Shengbang Tong and Jiachen Zhu and Koustuv Sinha and Zhuang Liu and Xinlei Chen and Michael Rabbat and Nicolas Ballas and Yann LeCun and Amir Bar and Saining Xie},
   year={2025},
   eprint={2504.01017},
@@ -315,7 +315,7 @@ webssl_dino2b_full2b = ModelMeta(
     use_instructions=False,
     training_datasets=webssl_dino_training_datasets,
     citation="""@article{fan2025scaling,
-  title={Scaling Language-Free Visual Representation Learning}, 
+  title={Scaling Language-Free Visual Representation Learning},
   author={David Fan and Shengbang Tong and Jiachen Zhu and Koustuv Sinha and Zhuang Liu and Xinlei Chen and Michael Rabbat and Nicolas Ballas and Yann LeCun and Amir Bar and Saining Xie},
   year={2025},
   eprint={2504.01017},
@@ -346,7 +346,7 @@ webssl_dino3b_full2b = ModelMeta(
     use_instructions=False,
     training_datasets=webssl_dino_training_datasets,
     citation="""@article{fan2025scaling,
-  title={Scaling Language-Free Visual Representation Learning}, 
+  title={Scaling Language-Free Visual Representation Learning},
   author={David Fan and Shengbang Tong and Jiachen Zhu and Koustuv Sinha and Zhuang Liu and Xinlei Chen and Michael Rabbat and Nicolas Ballas and Yann LeCun and Amir Bar and Saining Xie},
   year={2025},
   eprint={2504.01017},
@@ -377,7 +377,7 @@ webssl_dino5b_full2b = ModelMeta(
     use_instructions=False,
     training_datasets=webssl_dino_training_datasets,
     citation="""@article{fan2025scaling,
-  title={Scaling Language-Free Visual Representation Learning}, 
+  title={Scaling Language-Free Visual Representation Learning},
   author={David Fan and Shengbang Tong and Jiachen Zhu and Koustuv Sinha and Zhuang Liu and Xinlei Chen and Michael Rabbat and Nicolas Ballas and Yann LeCun and Amir Bar and Saining Xie},
   year={2025},
   eprint={2504.01017},
@@ -408,7 +408,7 @@ webssl_dino7b_full8b_224 = ModelMeta(
     use_instructions=False,
     training_datasets=webssl_dino_training_datasets,
     citation="""@article{fan2025scaling,
-  title={Scaling Language-Free Visual Representation Learning}, 
+  title={Scaling Language-Free Visual Representation Learning},
   author={David Fan and Shengbang Tong and Jiachen Zhu and Koustuv Sinha and Zhuang Liu and Xinlei Chen and Michael Rabbat and Nicolas Ballas and Yann LeCun and Amir Bar and Saining Xie},
   year={2025},
   eprint={2504.01017},
@@ -439,7 +439,7 @@ webssl_dino7b_full8b_378 = ModelMeta(
     use_instructions=False,
     training_datasets=webssl_dino_training_datasets,
     citation="""@article{fan2025scaling,
-  title={Scaling Language-Free Visual Representation Learning}, 
+  title={Scaling Language-Free Visual Representation Learning},
   author={David Fan and Shengbang Tong and Jiachen Zhu and Koustuv Sinha and Zhuang Liu and Xinlei Chen and Michael Rabbat and Nicolas Ballas and Yann LeCun and Amir Bar and Saining Xie},
   year={2025},
   eprint={2504.01017},
@@ -470,7 +470,7 @@ webssl_dino7b_full8b_518 = ModelMeta(
     use_instructions=False,
     training_datasets=webssl_dino_training_datasets,
     citation="""@article{fan2025scaling,
-  title={Scaling Language-Free Visual Representation Learning}, 
+  title={Scaling Language-Free Visual Representation Learning},
   author={David Fan and Shengbang Tong and Jiachen Zhu and Koustuv Sinha and Zhuang Liu and Xinlei Chen and Michael Rabbat and Nicolas Ballas and Yann LeCun and Amir Bar and Saining Xie},
   year={2025},
   eprint={2504.01017},
@@ -502,7 +502,7 @@ webssl_dino2b_light2b = ModelMeta(
     use_instructions=False,
     training_datasets=webssl_dino_training_datasets,
     citation="""@article{fan2025scaling,
-  title={Scaling Language-Free Visual Representation Learning}, 
+  title={Scaling Language-Free Visual Representation Learning},
   author={David Fan and Shengbang Tong and Jiachen Zhu and Koustuv Sinha and Zhuang Liu and Xinlei Chen and Michael Rabbat and Nicolas Ballas and Yann LeCun and Amir Bar and Saining Xie},
   year={2025},
   eprint={2504.01017},
@@ -533,7 +533,7 @@ webssl_dino2b_heavy2b = ModelMeta(
     use_instructions=False,
     training_datasets=webssl_dino_training_datasets,
     citation="""@article{fan2025scaling,
-  title={Scaling Language-Free Visual Representation Learning}, 
+  title={Scaling Language-Free Visual Representation Learning},
   author={David Fan and Shengbang Tong and Jiachen Zhu and Koustuv Sinha and Zhuang Liu and Xinlei Chen and Michael Rabbat and Nicolas Ballas and Yann LeCun and Amir Bar and Saining Xie},
   year={2025},
   eprint={2504.01017},
@@ -564,7 +564,7 @@ webssl_dino3b_light2b = ModelMeta(
     use_instructions=False,
     training_datasets=webssl_dino_training_datasets,
     citation="""@article{fan2025scaling,
-  title={Scaling Language-Free Visual Representation Learning}, 
+  title={Scaling Language-Free Visual Representation Learning},
   author={David Fan and Shengbang Tong and Jiachen Zhu and Koustuv Sinha and Zhuang Liu and Xinlei Chen and Michael Rabbat and Nicolas Ballas and Yann LeCun and Amir Bar and Saining Xie},
   year={2025},
   eprint={2504.01017},
@@ -595,7 +595,7 @@ webssl_dino3b_heavy2b = ModelMeta(
     use_instructions=False,
     training_datasets=webssl_dino_training_datasets,
     citation="""@article{fan2025scaling,
-  title={Scaling Language-Free Visual Representation Learning}, 
+  title={Scaling Language-Free Visual Representation Learning},
   author={David Fan and Shengbang Tong and Jiachen Zhu and Koustuv Sinha and Zhuang Liu and Xinlei Chen and Michael Rabbat and Nicolas Ballas and Yann LeCun and Amir Bar and Saining Xie},
   year={2025},
   eprint={2504.01017},
@@ -626,7 +626,7 @@ webssl_mae300m_full2b = ModelMeta(
     use_instructions=False,
     training_datasets=webssl_dino_training_datasets,
     citation="""@article{fan2025scaling,
-  title={Scaling Language-Free Visual Representation Learning}, 
+  title={Scaling Language-Free Visual Representation Learning},
   author={David Fan and Shengbang Tong and Jiachen Zhu and Koustuv Sinha and Zhuang Liu and Xinlei Chen and Michael Rabbat and Nicolas Ballas and Yann LeCun and Amir Bar and Saining Xie},
   year={2025},
   eprint={2504.01017},
@@ -657,7 +657,7 @@ webssl_mae700m_full2b = ModelMeta(
     use_instructions=False,
     training_datasets=webssl_dino_training_datasets,
     citation="""@article{fan2025scaling,
-  title={Scaling Language-Free Visual Representation Learning}, 
+  title={Scaling Language-Free Visual Representation Learning},
   author={David Fan and Shengbang Tong and Jiachen Zhu and Koustuv Sinha and Zhuang Liu and Xinlei Chen and Michael Rabbat and Nicolas Ballas and Yann LeCun and Amir Bar and Saining Xie},
   year={2025},
   eprint={2504.01017},
@@ -688,7 +688,7 @@ webssl_mae1b_full2b = ModelMeta(
     use_instructions=False,
     training_datasets=webssl_dino_training_datasets,
     citation="""@article{fan2025scaling,
-  title={Scaling Language-Free Visual Representation Learning}, 
+  title={Scaling Language-Free Visual Representation Learning},
   author={David Fan and Shengbang Tong and Jiachen Zhu and Koustuv Sinha and Zhuang Liu and Xinlei Chen and Michael Rabbat and Nicolas Ballas and Yann LeCun and Amir Bar and Saining Xie},
   year={2025},
   eprint={2504.01017},

mteb/models/model_implementations/emillykkejensen_models.py

@@ -2,7 +2,7 @@ from mteb.models.model_meta import ModelMeta
 from mteb.models.sentence_transformer_wrapper import sentence_transformers_loader
 
 embedding_gemma_300m_scandi = ModelMeta(
-    loader=sentence_transformers_loader,  # type: ignore
+    loader=sentence_transformers_loader,
     name="emillykkejensen/EmbeddingGemma-Scandi-300m",
     model_type=["dense"],
     languages=["dan-Latn", "swe-Latn", "nor-Latn", "nob-Latn", "nno-Latn"],
@@ -35,7 +35,7 @@ embedding_gemma_300m_scandi = ModelMeta(
 
 
 qwen_scandi = ModelMeta(
-    loader=sentence_transformers_loader,  # type: ignore
+    loader=sentence_transformers_loader,
     name="emillykkejensen/Qwen3-Embedding-Scandi-0.6B",
     model_type=["dense"],
     languages=["dan-Latn", "swe-Latn", "nor-Latn", "nob-Latn", "nno-Latn"],
@@ -59,7 +59,7 @@ qwen_scandi = ModelMeta(
 
 
 mmbert_scandi = ModelMeta(
-    loader=sentence_transformers_loader,  # type: ignore
+    loader=sentence_transformers_loader,
     name="emillykkejensen/mmBERTscandi-base-embedding",
     model_type=["dense"],
     languages=["dan-Latn", "swe-Latn", "nor-Latn", "nob-Latn", "nno-Latn"],

mteb/models/model_implementations/gme_v_models.py

@@ -2,6 +2,7 @@ from __future__ import annotations
 
 import logging
 import math
+import warnings
 from typing import TYPE_CHECKING, Any
 
 import torch
@@ -261,9 +262,9 @@ def smart_resize(
         w_bar = ceil_by_factor(width * beta, factor)
 
     if max(h_bar, w_bar) / min(h_bar, w_bar) > MAX_RATIO:
-        logger.warning(
-            f"Absolute aspect ratio must be smaller than {MAX_RATIO}, got {max(h_bar, w_bar) / min(h_bar, w_bar)}"
-        )
+        msg = f"Absolute aspect ratio must be smaller than {MAX_RATIO}, got {max(h_bar, w_bar) / min(h_bar, w_bar)}"
+        logger.warning(msg)
+        warnings.warn(msg)
         if h_bar > w_bar:
             h_bar = w_bar * MAX_RATIO
         else:

mteb/models/model_implementations/jina_clip.py

@@ -121,7 +121,7 @@ class JinaCLIPModel(AbsEncoder):
 
 
 jina_clip_v1 = ModelMeta(
-    loader=JinaCLIPModel,  # type: ignore
+    loader=JinaCLIPModel,
     name="jinaai/jina-clip-v1",
     model_type=["dense"],
     languages=["eng-Latn"],

mteb/models/model_implementations/jina_models.py

@@ -795,7 +795,7 @@ jina_embeddings_v4 = ModelMeta(
 
 
 jina_embeddings_v3 = ModelMeta(
-    loader=JinaWrapper,  # type: ignore
+    loader=JinaWrapper,
     loader_kwargs=dict(
         trust_remote_code=True,
         model_prompts={