mteb 2.5.2__py3-none-any.whl → 2.7.2__py3-none-any.whl

mteb/models/model_implementations/ru_sentence_models.py

@@ -250,7 +250,7 @@ rubert_tiny = ModelMeta(
     max_tokens=512,
     reference="https://huggingface.co/cointegrated/rubert-tiny",
     similarity_fn_name=ScoringFunction.COSINE,
-    framework=["Sentence Transformers", "PyTorch"],
+    framework=["Sentence Transformers", "PyTorch", "Transformers", "safetensors"],
     use_instructions=False,
     public_training_code="https://gist.github.com/avidale/7bc6350f26196918bf339c01261f5c60",
     training_datasets={
@@ -276,7 +276,7 @@ rubert_tiny2 = ModelMeta(
     max_tokens=2048,
     reference="https://huggingface.co/cointegrated/rubert-tiny2",
     similarity_fn_name=ScoringFunction.COSINE,
-    framework=["Sentence Transformers", "PyTorch"],
+    framework=["Sentence Transformers", "PyTorch", "safetensors", "Transformers"],
     use_instructions=False,
     public_training_code="https://colab.research.google.com/drive/1mSWfIQ6PIlteLVZ9DKKpcorycgLIKZLf?usp=sharing",
     training_datasets=set(
@@ -303,7 +303,7 @@ sbert_large_nlu_ru = ModelMeta(
     max_tokens=512,  # best guess
     reference="https://huggingface.co/ai-forever/sbert_large_nlu_ru",
     similarity_fn_name=ScoringFunction.COSINE,
-    framework=["Sentence Transformers", "PyTorch"],
+    framework=["Sentence Transformers", "PyTorch", "Transformers", "safetensors"],
     use_instructions=False,
     public_training_code=None,
     public_training_data=None,
@@ -329,7 +329,7 @@ sbert_large_mt_nlu_ru = ModelMeta(
     max_tokens=512,  # best guess
     reference="https://huggingface.co/ai-forever/sbert_large_mt_nlu_ru",
     similarity_fn_name=ScoringFunction.COSINE,
-    framework=["Sentence Transformers", "PyTorch"],
+    framework=["Sentence Transformers", "PyTorch", "Transformers", "safetensors"],
     use_instructions=False,
     public_training_code=None,
     public_training_data=None,
@@ -357,7 +357,7 @@ user_base_ru = ModelMeta(
     max_tokens=512,
     reference="https://huggingface.co/deepvk/USER-base",
     similarity_fn_name=ScoringFunction.COSINE,
-    framework=["Sentence Transformers", "PyTorch"],
+    framework=["Sentence Transformers", "PyTorch", "safetensors"],
     adapted_from="https://huggingface.co/deepvk/deberta-v1-base",
     use_instructions=True,
     citation="""@misc{deepvk2024user,
@@ -418,7 +418,7 @@ user_bge_m3 = ModelMeta(
     max_tokens=8194,
     reference="https://huggingface.co/deepvk/USER-base",
     similarity_fn_name=ScoringFunction.COSINE,
-    framework=["Sentence Transformers", "PyTorch"],
+    framework=["Sentence Transformers", "PyTorch", "safetensors"],
     adapted_from="BAAI/bge-m3",
     use_instructions=False,
     training_datasets={
@@ -469,7 +469,7 @@ deberta_v1_ru = ModelMeta(
     max_tokens=512,
     reference="https://huggingface.co/deepvk/deberta-v1-base",
     similarity_fn_name=ScoringFunction.COSINE,
-    framework=["Sentence Transformers", "PyTorch"],
+    framework=["Sentence Transformers", "PyTorch", "Transformers", "safetensors"],
     use_instructions=False,
     # Wikipedia, Books, Twitter comments, Pikabu, Proza.ru, Film subtitles, News websites, and Social corpus
     public_training_code=None,
@@ -500,7 +500,7 @@ rubert_base_cased = ModelMeta(
     max_tokens=512,
     reference="https://huggingface.co/DeepPavlov/rubert-base-cased",
     similarity_fn_name=ScoringFunction.COSINE,
-    framework=["Sentence Transformers", "PyTorch"],
+    framework=["Sentence Transformers", "PyTorch", "Transformers"],
     use_instructions=False,
     public_training_code=None,
     public_training_data=None,
@@ -536,7 +536,7 @@ distilrubert_small_cased_conversational = ModelMeta(
     max_tokens=512,
     reference="https://huggingface.co/DeepPavlov/distilrubert-small-cased-conversational",
     similarity_fn_name=ScoringFunction.COSINE,
-    framework=["Sentence Transformers", "PyTorch"],
+    framework=["Sentence Transformers", "PyTorch", "Transformers"],
     use_instructions=False,
     public_training_code=None,
     public_training_data=None,
@@ -571,7 +571,7 @@ rubert_base_cased_sentence = ModelMeta(
     max_tokens=512,
     reference="https://huggingface.co/DeepPavlov/rubert-base-cased-sentence",
     similarity_fn_name=ScoringFunction.COSINE,
-    framework=["Sentence Transformers", "PyTorch"],
+    framework=["Sentence Transformers", "PyTorch", "Transformers"],
     use_instructions=False,
     public_training_code=None,
     public_training_data=None,
@@ -596,7 +596,7 @@ labse_en_ru = ModelMeta(
     max_tokens=512,
     reference="https://huggingface.co/cointegrated/LaBSE-en-ru",
     similarity_fn_name=ScoringFunction.COSINE,
-    framework=["Sentence Transformers", "PyTorch"],
+    framework=["Sentence Transformers", "PyTorch", "Transformers", "safetensors"],
     use_instructions=False,
     public_training_code="https://colab.research.google.com/drive/1dnPRn0-ugj3vZgSpyCC9sgslM2SuSfHy?usp=sharing",
     public_training_data=None,
@@ -624,7 +624,7 @@ rubert_tiny_turbo = ModelMeta(
     max_tokens=2048,
     reference="https://huggingface.co/sergeyzh/rubert-tiny-turbo",
     similarity_fn_name=ScoringFunction.COSINE,
-    framework=["Sentence Transformers", "PyTorch"],
+    framework=["Sentence Transformers", "PyTorch", "safetensors", "Transformers"],
     use_instructions=False,
     public_training_code=None,
     public_training_data=None,
@@ -647,7 +647,7 @@ rubert_mini_frida = ModelMeta(
     max_tokens=2048,
     reference="https://huggingface.co/sergeyzh/rubert-mini-frida",
     similarity_fn_name=ScoringFunction.COSINE,
-    framework=["Sentence Transformers", "PyTorch"],
+    framework=["Sentence Transformers", "PyTorch", "safetensors", "Transformers"],
     use_instructions=True,
     public_training_code=None,
     public_training_data=None,
@@ -675,7 +675,7 @@ labse_ru_turbo = ModelMeta(
     max_tokens=512,
     reference="https://huggingface.co/sergeyzh/LaBSE-ru-turbo",
     similarity_fn_name=ScoringFunction.COSINE,
-    framework=["Sentence Transformers", "PyTorch"],
+    framework=["Sentence Transformers", "PyTorch", "safetensors", "Transformers"],
     use_instructions=False,
     training_datasets=turbo_models_datasets,
     public_training_code=None,
@@ -745,7 +745,7 @@ rosberta_ru_en = ModelMeta(
     },
     public_training_data=None,
     public_training_code=None,
-    framework=["Sentence Transformers", "PyTorch"],
+    framework=["Sentence Transformers", "PyTorch", "safetensors", "Transformers"],
     citation="""@misc{snegirev2024russianfocusedembeddersexplorationrumteb,
       title={The Russian-focused embedders' exploration: ruMTEB benchmark and Russian embedding model design},
       author={Artem Snegirev and Maria Tikhonova and Anna Maksimova and Alena Fenogenova and Alexander Abramov},
@@ -895,7 +895,7 @@ frida = ModelMeta(
     training_datasets=frida_training_datasets,
     public_training_data=None,
     public_training_code=None,
-    framework=["Sentence Transformers", "PyTorch"],
+    framework=["Sentence Transformers", "PyTorch", "safetensors", "Transformers"],
     citation=None,
 )
 
@@ -924,7 +924,7 @@ giga_embeddings = ModelMeta(
     max_tokens=4096,
     reference="https://huggingface.co/ai-sage/Giga-Embeddings-instruct",
     similarity_fn_name=ScoringFunction.COSINE,
-    framework=["Sentence Transformers", "PyTorch"],
+    framework=["Sentence Transformers", "PyTorch", "safetensors", "Transformers"],
     use_instructions=True,
     public_training_code=None,
     public_training_data=None,
@@ -956,7 +956,7 @@ berta = ModelMeta(
     max_tokens=512,
     reference="https://huggingface.co/sergeyzh/BERTA",
     similarity_fn_name=ScoringFunction.COSINE,
-    framework=["Sentence Transformers", "PyTorch"],
+    framework=["Sentence Transformers", "PyTorch", "safetensors", "Transformers"],
     use_instructions=True,
     training_datasets=berta_training_datasets,
     public_training_code=None,
@@ -1034,7 +1034,7 @@ user2_small = ModelMeta(
     training_datasets=user2_training_data,
     public_training_data=None,
     public_training_code="https://github.com/BlessedTatonka/some_code/tree/2899f27d51efdf4217fc6453799ff197e9792f1e",
-    framework=["Sentence Transformers", "PyTorch"],
+    framework=["Sentence Transformers", "PyTorch", "safetensors"],
     citation="""@misc{deepvk2025user,
     title={USER2},
     author={Malashenko, Boris and Spirin, Egor and Sokolov Andrey},
@@ -1067,7 +1067,7 @@ user2_base = ModelMeta(
     training_datasets=user2_training_data,
     public_training_data=None,
     public_training_code="https://github.com/BlessedTatonka/some_code/tree/2899f27d51efdf4217fc6453799ff197e9792f1e",
-    framework=["Sentence Transformers", "PyTorch"],
+    framework=["Sentence Transformers", "PyTorch", "safetensors"],
     citation="""@misc{deepvk2025user,
     title={USER2},
     author={Malashenko, Boris and Spirin, Egor and Sokolov Andrey},

mteb/models/model_implementations/ruri_models.py

@@ -44,7 +44,7 @@ cl_nagoya_ruri_v3_30m = ModelMeta(
     max_tokens=8192,
     reference="https://huggingface.co/cl-nagoya/ruri-v3-30m",
     similarity_fn_name="cosine",
-    framework=["PyTorch", "Sentence Transformers"],
+    framework=["PyTorch", "Sentence Transformers", "safetensors"],
     use_instructions=True,
     superseded_by=None,
     training_datasets={
@@ -75,7 +75,7 @@ cl_nagoya_ruri_v3_70m = ModelMeta(
     max_tokens=8192,
     reference="https://huggingface.co/cl-nagoya/ruri-v3-70m",
     similarity_fn_name="cosine",
-    framework=["PyTorch", "Sentence Transformers"],
+    framework=["PyTorch", "Sentence Transformers", "safetensors"],
     use_instructions=True,
     superseded_by=None,
     training_datasets={"MrTidyRetrieval", "MIRACLRetrieval"},
@@ -104,7 +104,7 @@ cl_nagoya_ruri_v3_130m = ModelMeta(
     max_tokens=8192,
     reference="https://huggingface.co/cl-nagoya/ruri-v3-130m",
     similarity_fn_name="cosine",
-    framework=["PyTorch", "Sentence Transformers"],
+    framework=["PyTorch", "Sentence Transformers", "safetensors"],
     use_instructions=True,
     superseded_by=None,
     training_datasets={"MrTidyRetrieval", "MIRACLRetrieval"},
@@ -133,7 +133,7 @@ cl_nagoya_ruri_v3_310m = ModelMeta(
     max_tokens=8192,
     reference="https://huggingface.co/cl-nagoya/ruri-v3-310m",
     similarity_fn_name="cosine",
-    framework=["PyTorch", "Sentence Transformers"],
+    framework=["PyTorch", "Sentence Transformers", "safetensors"],
     use_instructions=True,
     superseded_by=None,
     training_datasets={"MrTidyRetrieval", "MIRACLRetrieval"},
@@ -163,7 +163,7 @@ cl_nagoya_ruri_small_v2 = ModelMeta(
     max_tokens=512,
     reference="https://huggingface.co/cl-nagoya/ruri-small-v2",
     similarity_fn_name="cosine",
-    framework=["PyTorch", "Sentence Transformers"],
+    framework=["PyTorch", "Sentence Transformers", "safetensors"],
     use_instructions=True,
     adapted_from="line-corporation/line-distilbert-base-japanese",
     superseded_by=None,
@@ -192,7 +192,7 @@ cl_nagoya_ruri_base_v2 = ModelMeta(
     max_tokens=512,
     reference="https://huggingface.co/cl-nagoya/ruri-base-v2",
     similarity_fn_name="cosine",
-    framework=["PyTorch", "Sentence Transformers"],
+    framework=["PyTorch", "Sentence Transformers", "safetensors"],
     use_instructions=True,
     adapted_from="tohoku-nlp/bert-base-japanese-v3",
     superseded_by=None,
@@ -221,7 +221,7 @@ cl_nagoya_ruri_large_v2 = ModelMeta(
     max_tokens=512,
     reference="https://huggingface.co/cl-nagoya/ruri-large-v2",
     similarity_fn_name="cosine",
-    framework=["PyTorch", "Sentence Transformers"],
+    framework=["PyTorch", "Sentence Transformers", "safetensors"],
     use_instructions=True,
     adapted_from="tohoku-nlp/bert-large-japanese-v2",
     superseded_by=None,
@@ -251,7 +251,7 @@ cl_nagoya_ruri_small_v1 = ModelMeta(
     max_tokens=512,
     reference="https://huggingface.co/cl-nagoya/ruri-small",
     similarity_fn_name="cosine",
-    framework=["PyTorch", "Sentence Transformers"],
+    framework=["PyTorch", "Sentence Transformers", "safetensors"],
     use_instructions=True,
     adapted_from="line-corporation/line-distilbert-base-japanese",
     superseded_by="cl-nagoya/ruri-small-v2",
@@ -280,7 +280,7 @@ cl_nagoya_ruri_base_v1 = ModelMeta(
     max_tokens=512,
     reference="https://huggingface.co/cl-nagoya/ruri-base",
     similarity_fn_name="cosine",
-    framework=["PyTorch", "Sentence Transformers"],
+    framework=["PyTorch", "Sentence Transformers", "safetensors"],
     use_instructions=True,
     adapted_from="tohoku-nlp/bert-base-japanese-v3",
     superseded_by="cl-nagoya/ruri-base-v2",
@@ -310,7 +310,7 @@ cl_nagoya_ruri_large_v1 = ModelMeta(
     max_tokens=512,
     reference="https://huggingface.co/cl-nagoya/ruri-large",
     similarity_fn_name="cosine",
-    framework=["PyTorch", "Sentence Transformers"],
+    framework=["PyTorch", "Sentence Transformers", "safetensors"],
     use_instructions=True,
     adapted_from="tohoku-nlp/bert-large-japanese-v2",
     superseded_by="cl-nagoya/ruri-large-v2",

mteb/models/model_implementations/salesforce_models.py

@@ -58,7 +58,7 @@ SFR_Embedding_2_R = ModelMeta(
     max_tokens=32768,
     reference="https://huggingface.co/Salesforce/SFR-Embedding-2_R",
     similarity_fn_name=ScoringFunction.COSINE,
-    framework=["Sentence Transformers", "PyTorch"],
+    framework=["Sentence Transformers", "PyTorch", "safetensors", "Transformers"],
     use_instructions=True,
     adapted_from="intfloat/e5-mistral-7b-instruct",
     public_training_code=None,
@@ -96,7 +96,7 @@ SFR_Embedding_Code_2B_R = ModelMeta(
     max_tokens=8192,
     reference="https://huggingface.co/Salesforce/SFR-Embedding-Code-2B_R",
     similarity_fn_name=ScoringFunction.COSINE,
-    framework=["Sentence Transformers", "PyTorch"],
+    framework=["Sentence Transformers", "PyTorch", "safetensors", "Transformers"],
     use_instructions=True,
     adapted_from="google/gemma-2-2b-it",
     public_training_code=None,
@@ -134,7 +134,7 @@ SFR_Embedding_Mistral = ModelMeta(
     max_tokens=32768,
     reference="https://huggingface.co/Salesforce/SFR-Embedding-Mistral",
     similarity_fn_name=ScoringFunction.COSINE,
-    framework=["Sentence Transformers", "PyTorch"],
+    framework=["Sentence Transformers", "PyTorch", "safetensors", "Transformers"],
     use_instructions=True,
     public_training_code=None,
     public_training_data=None,

mteb/models/model_implementations/samilpwc_models.py

@@ -57,7 +57,7 @@ samilpwc_expr = ModelMeta(
     max_tokens=514,
     reference="https://huggingface.co/SamilPwC-AXNode-GenAI/PwC-Embedding_expr",
     similarity_fn_name="cosine",
-    framework=["Sentence Transformers", "PyTorch"],
+    framework=["Sentence Transformers", "PyTorch", "safetensors", "Transformers"],
     use_instructions=True,
     public_training_code=None,
     public_training_data=None,

mteb/models/model_implementations/sarashina_embedding_models.py

@@ -130,7 +130,7 @@ sbintuitions_sarashina_embedding_v2_1b = ModelMeta(
     max_tokens=8192,
     reference="https://huggingface.co/sbintuitions/sarashina-embedding-v2-1b",
     similarity_fn_name="cosine",
-    framework=["Sentence Transformers", "PyTorch"],
+    framework=["Sentence Transformers", "PyTorch", "safetensors", "Transformers"],
     use_instructions=True,
     adapted_from="sbintuitions/sarashina2.2-1b",
     superseded_by=None,
@@ -156,7 +156,7 @@ sbintuitions_sarashina_embedding_v1_1b = ModelMeta(
     max_tokens=8192,
     reference="https://huggingface.co/sbintuitions/sarashina-embedding-v1-1b",
     similarity_fn_name="cosine",
-    framework=["Sentence Transformers", "PyTorch"],
+    framework=["Sentence Transformers", "PyTorch", "safetensors", "Transformers"],
     use_instructions=False,
     adapted_from="sbintuitions/sarashina2.1-1b",
     superseded_by="sbintuitions/sarashina-embedding-v2-1b",

mteb/models/model_implementations/searchmap_models.py

@@ -33,7 +33,7 @@ searchmap_preview = ModelMeta(
     max_tokens=8192,
     reference="https://huggingface.co/VPLabs/SearchMap_Preview",
     similarity_fn_name="cosine",
-    framework=["Sentence Transformers", "PyTorch"],
+    framework=["Sentence Transformers", "PyTorch", "Transformers", "safetensors"],
     public_training_code=None,
     public_training_data=None,
     training_datasets=None,

mteb/models/model_implementations/seed_1_6_embedding_models_1215.py

@@ -4,13 +4,15 @@ import base64
 import logging
 import os
 import time
-from concurrent.futures import ThreadPoolExecutor, as_completed
+from concurrent.futures import ThreadPoolExecutor
+from functools import partial
 from io import BytesIO
 from typing import TYPE_CHECKING, Any
 
 import requests
 import torch
 from torch.utils.data import DataLoader
+from tqdm import tqdm
 
 from mteb._requires_package import requires_package
 from mteb.abstasks.task_metadata import TaskMetadata
@@ -26,114 +28,6 @@ if TYPE_CHECKING:
 
 logger = logging.getLogger(__name__)
 
-
-def pil_to_base64(image, format="jpeg"):
-    if image is None:
-        return None
-    buffer = BytesIO()
-    image.save(buffer, format=format)
-    img_bytes = buffer.getvalue()
-    encoded_bytes = base64.b64encode(img_bytes)
-    return encoded_bytes.decode("utf-8")
-
-
-def multimodal_embedding(image_base64=None, text_content=None):
-    auth_token = os.getenv("VOLCES_AUTH_TOKEN")
-    model_name = "doubao-embedding-vision-251215"
-    api_url = "https://ark.cn-beijing.volces.com/api/v3/embeddings/multimodal"
-
-    headers = {
-        "Authorization": f"Bearer {auth_token}",
-        "x-ark-vlm1": "true",
-        "Content-Type": "application/json",
-    }
-
-    if image_base64 is not None and text_content is None:
-        inputs = []
-        for image in image_base64:
-            image_format = "jpeg"
-            image_data = f"data:image/{image_format};base64,{image}"
-            inputs.append({"type": "image_url", "image_url": {"url": image_data}})
-
-        payload = {"model": model_name, "input": inputs}
-    elif image_base64 is None and text_content is not None:
-        payload = {
-            "model": model_name,
-            "input": [
-                {"type": "text", "text": text_content},
-            ],
-        }
-    else:
-        inputs = []
-        for image in image_base64:
-            image_format = "jpeg"
-            image_data = f"data:image/{image_format};base64,{image}"
-            inputs.append({"type": "image_url", "image_url": {"url": image_data}})
-        inputs.append({"type": "text", "text": text_content})
-        payload = {"model": model_name, "input": inputs}
-
-    try:
-        response = requests.post(url=api_url, headers=headers, json=payload, timeout=10)
-
-        response.raise_for_status()
-        return response.json()
-
-    except requests.exceptions.HTTPError as http_err:
-        logger.error(f"HTTP error ({http_err.response.status_code}): {http_err}")
-    except requests.exceptions.JSONDecodeError:
-        logger.error("Error:The response is not in valid JSON format")
-    except requests.exceptions.Timeout:
-        logger.error("Error:Request timeout")
-    except Exception as e:
-        logger.error(f"Unknown error: {str(e)}")
-
-    return None
-
-
-def multi_thread_encode(sentences, batch_size=1, max_workers=8):
-    batches = []
-    for idx in range(0, len(sentences), batch_size):
-        batches.append((idx // batch_size, sentences[idx : idx + batch_size]))
-
-    n_batches = len(batches)
-    results = [None] * n_batches  # Pre-allocated result list
-    all_embeddings = []  # Final ordered embeddings
-
-    def _process_batch(batch_idx, batch_sentences):
-        sentence = batch_sentences[0]
-
-        retries = 5
-        while retries > 0:
-            try:
-                resp = multimodal_embedding(text_content=sentence)
-                embedding = torch.tensor(resp["data"]["embedding"])
-                break
-            except Exception as e:
-                time.sleep(1)
-                logger.warning(f"Retrying... {retries} retries left. Error: {str(e)}")
-                retries -= 1
-                if retries == 0:
-                    raise e
-        return batch_idx, embedding
-
-    with ThreadPoolExecutor(max_workers=max_workers) as executor:
-        futures = {
-            executor.submit(_process_batch, idx, batch): idx for idx, batch in batches
-        }
-
-        for future in as_completed(futures):
-            batch_idx, embeddings = future.result()
-            results[batch_idx] = embeddings
-
-    for batch_embeddings in results:
-        all_embeddings.append(batch_embeddings)
-
-    all_embeddings = torch.stack(all_embeddings, dim=0)
-    all_embeddings = torch.nn.functional.normalize(all_embeddings, dim=-1)
-
-    return all_embeddings.float().cpu()
-
-
 doubao_embedding_training_data = (
     {
         "PawsXPairClassification",
@@ -166,25 +60,80 @@ class Seed16EmbeddingWrapper(AbsEncoder):
             "pip install mteb[ark]",
             "tiktoken",
         )
-        import tiktoken
 
         self._model_name = model_name
         self._max_tokens = 32768
         self._embed_dim = embed_dim
         self._available_embed_dims = [2048, 1024]
-        self._encoding = tiktoken.get_encoding(tokenizer_name)
 
-    def truncate_text_tokens(self, text: str) -> str:
-        """Truncate a string to have `max_tokens` according to the given encoding.
+    def pil_to_base64(self, image, format="jpeg"):
+        if image is None:
+            return None
+        buffer = BytesIO()
+        image.save(buffer, format=format)
+        img_bytes = buffer.getvalue()
+        encoded_bytes = base64.b64encode(img_bytes)
+        return encoded_bytes.decode("utf-8")
+
+    def multimodal_embedding(self, instruction, image_base64, text_content):
+        auth_token = os.getenv("VOLCES_AUTH_TOKEN")
+        model_name = "doubao-embedding-vision-251215"
+        api_url = "https://ark.cn-beijing.volces.com/api/v3/embeddings/multimodal"
+
+        headers = {
+            "Authorization": f"Bearer {auth_token}",
+            "x-ark-vlm1": "true",
+            "Content-Type": "application/json",
+        }
 
-        Args:
-            text: The input string to be truncated.
+        if text_content is not None and len(text_content) > self._max_tokens:
+            text_content = text_content[: self._max_tokens]
+
+        if image_base64 is not None and text_content is None:
+            inputs = []
+            for image in image_base64:
+                image_format = "jpeg"
+                image_data = f"data:image/{image_format};base64,{image}"
+                inputs.append({"type": "image_url", "image_url": {"url": image_data}})
+
+            payload = {"model": model_name, "input": inputs}
+        elif image_base64 is None and text_content is not None:
+            payload = {
+                "model": model_name,
+                "instruction": instruction,
+                "input": [
+                    {"type": "text", "text": text_content},
+                ],
+            }
+        else:
+            inputs = []
+            for image in image_base64:
+                image_format = "jpeg"
+                image_data = f"data:image/{image_format};base64,{image}"
+                inputs.append({"type": "image_url", "image_url": {"url": image_data}})
+            inputs.append({"type": "text", "text": text_content})
+            payload = {"model": model_name, "input": inputs}
+
+        max_retries = 3
+        retry_count = 0
+
+        while retry_count < max_retries:
+            response = requests.post(
+                url=api_url, headers=headers, json=payload, timeout=30
+            )
 
-        Returns:
-            The truncated string.
-        """
-        truncated_sentence = self._encoding.encode(text)[: self._max_tokens]
-        return self._encoding.decode(truncated_sentence)
+            if response.status_code != 200:
+                retry_count += 1
+                time.sleep(3)
+                continue
+
+            response_json = response.json()
+            return response_json
+
+        raise Exception(
+            f"Request failed with status code {response.status_code}. "
+            f"Response: {response.text}"
+        )
 
     def get_fused_embeddings(
         self,
@@ -204,59 +153,69 @@ class Seed16EmbeddingWrapper(AbsEncoder):
         if images is not None and texts is not None:
             assert len(texts) == len(images)
             batch_len = len(texts)
-            images_base64 = [pil_to_base64(image) for image in images]
+            images_base64 = [self.pil_to_base64(image) for image in images]
         elif images is None:
             batch_len = len(texts)
             images_base64 = [None for _ in range(batch_len)]
         elif texts is None:
             batch_len = len(images)
-            images_base64 = [pil_to_base64(image) for image in images]
+            images_base64 = [self.pil_to_base64(image) for image in images]
         else:
             raise ValueError("images and texts cannot be None at the same time")
 
-        outputs = []
-        for i in range(batch_len):
+        def process_item(
+            i, prompt_type, task_name, texts, images_base64, multimodal_embedding
+        ):
             if (
                 prompt_type == PromptType("query") or prompt_type is None
             ) and task_name in TASK_NAME_TO_INSTRUCTION:
                 instruction = TASK_NAME_TO_INSTRUCTION[task_name]
                 instruction = instruction.rstrip("{}").rstrip("\n")
-                if texts[i] != "":
-                    input_text = (
-                        "Target_modality:Text.\n Instruction:"
-                        + instruction
-                        + "\n Query:{}"
-                    ).format(texts[i])
-                else:
-                    input_text = (
-                        "Target_modality:Text.\n Instruction:"
-                        + instruction
-                        + "\n Query:"
-                    )
+                instruction = (
+                    "Target_modality:Text.\n Instruction:" + instruction + "\n Query:"
+                )
+                input_text = texts[i]
             else:
                 if texts[i] != "" and images_base64[i] is not None:
-                    instruction = "Instruction: Compress the the text and image into one word.\n Query: {}"
-                    input_text = instruction.format(texts[i])
+                    instruction = "Instruction: Compress the text and image into one word.\n Query:"
+                    input_text = texts[i]
                 elif texts[i] != "":
                     instruction = (
-                        "Instruction: Compress the the text into one word.\n Query: {}"
+                        "Instruction: Compress the text into one word.\n Query:"
                     )
-                    input_text = instruction.format(texts[i])
+                    input_text = texts[i]
                 elif images_base64[i] is not None:
                     instruction = (
-                        "Instruction: Compress the the image into one word.\n Query:"
+                        "Instruction: Compress the image into one word.\n Query:"
                     )
-                    input_text = instruction
+                    input_text = None
                 else:
                     raise ValueError("image and text are both None")
 
             resp = multimodal_embedding(
-                image_base64=[images_base64[i]], text_content=input_text
+                instruction=instruction,
+                image_base64=images_base64[i],
+                text_content=input_text,
             )
             embedding = torch.tensor(resp["data"]["embedding"])
             embedding = torch.reshape(embedding, (1, -1))
+            return embedding
+
+        outputs = []
+        process_partial = partial(
+            process_item,
+            prompt_type=prompt_type,
+            task_name=task_name,
+            texts=texts,
+            images_base64=images_base64,
+            multimodal_embedding=self.multimodal_embedding,
+        )
+        with ThreadPoolExecutor(max_workers=15) as executor:
+            futures = [executor.submit(process_partial, i) for i in range(batch_len)]
+            for future in tqdm(futures, total=batch_len, desc="Encoding"):
+                outputs.append(future.result())
 
-        outputs = torch.stack(outputs, dim=0)
+        outputs = torch.stack(outputs, dim=0).squeeze(1)
 
         if self._embed_dim is not None:
             outputs = outputs[:, : self._embed_dim]
@@ -273,13 +232,21 @@ class Seed16EmbeddingWrapper(AbsEncoder):
         prompt_type: PromptType | None = None,
         **kwargs: Any,
     ) -> Array:
-        sentences = [text for batch in inputs for text in batch["text"]]
-        images = [image for batch in inputs for image in batch["image"]]
+        if "text" in inputs.dataset.features:
+            sentences = [text for batch in inputs for text in batch["text"]]
+        else:
+            sentences = None
+
+        if "image" in inputs.dataset.features:
+            images = [image for batch in inputs for image in batch["image"]]
+        else:
+            images = None
 
         return self.get_fused_embeddings(
             texts=sentences,
             images=images,
             task_name=task_metadata.name,
+            prompt_type=prompt_type,
             **kwargs,
         )