model: Add Qwen3 Embedding model

mteb/models/overview.py

@@ -68,6 +68,7 @@
     promptriever_models,
     qodo_models,
     qtack_models,
+    qwen3_models,
     repllama_models,
     rerankers_custom,
     rerankers_monot5_based,
@@ -143,6 +144,7 @@
     promptriever_models,
     qodo_models,
     qtack_models,
+    qwen3_models,
     repllama_models,
     rerankers_custom,
     rerankers_monot5_based,

mteb/models/qwen3_models.py

@@ -0,0 +1,198 @@
+from __future__ import annotations
+
+import os
+from functools import partial
+
+from mteb.encoder_interface import PromptType
+from mteb.model_meta import ModelMeta
+from mteb.models.instruct_wrapper import InstructSentenceTransformerWrapper
+
+
+def instruction_template(
+    instruction: str, prompt_type: PromptType | None = None
+) -> str:
+    if not instruction or prompt_type == PromptType.passage:
+        return ""
+    if isinstance(instruction, dict):
+        if prompt_type is None:
+            instruction = list(instruction.values())[0]  # TODO
+        else:
+            instruction = instruction[prompt_type]
+    return f"Instruct: {instruction}\nQuery:"
+
+
+multilingual_langs = [
+    "afr-Latn",
+    "ara-Arab",
+    "aze-Latn",
+    "bel-Cyrl",
+    "bul-Cyrl",
+    "ben-Beng",
+    "cat-Latn",
+    "ceb-Latn",
+    "ces-Latn",
+    "cym-Latn",
+    "dan-Latn",
+    "deu-Latn",
+    "ell-Grek",
+    "eng-Latn",
+    "spa-Latn",
+    "est-Latn",
+    "eus-Latn",
+    "fas-Arab",
+    "fin-Latn",
+    "fra-Latn",
+    "glg-Latn",
+    "guj-Gujr",
+    "heb-Hebr",
+    "hin-Deva",
+    "hrv-Latn",
+    "hat-Latn",
+    "hun-Latn",
+    "hye-Armn",
+    "ind-Latn",
+    "isl-Latn",
+    "ita-Latn",
+    "jpn-Jpan",
+    "jav-Latn",
+    "kat-Geor",
+    "kaz-Cyrl",
+    "khm-Khmr",
+    "kan-Knda",
+    "kor-Hang",
+    "kir-Cyrl",
+    "lao-Laoo",
+    "lit-Latn",
+    "lav-Latn",
+    "mkd-Cyrl",
+    "mal-Mlym",
+    "mon-Cyrl",
+    "mar-Deva",
+    "msa-Latn",
+    "mya-Mymr",
+    "nep-Deva",
+    "nld-Latn",
+    "nor-Latn",
+    "pan-Guru",
+    "pol-Latn",
+    "por-Latn",
+    "que-Latn",
+    "ron-Latn",
+    "rus-Cyrl",
+    "sin-Sinh",
+    "slk-Latn",
+    "slv-Latn",
+    "swa-Latn",
+    "tam-Taml",
+    "tel-Telu",
+    "tha-Thai",
+    "tgl-Latn",
+    "tur-Latn",
+    "ukr-Cyrl",
+    "urd-Arab",
+    "vie-Latn",
+    "yor-Latn",
+    "zho-Hans",
+]
+
+training_data = {
+    "T2Retrieval": ["train"],
+    "DuRetrieval": ["train"],
+    "MMarcoReranking": ["train"],
+    "CMedQAv2-reranking": ["train"],
+    "NQ": ["train"],
+    "MSMARCO": ["train"],
+    "HotpotQA": ["train"],
+    "FEVER": ["train"],
+    "MrTidyRetrieval": ["train"],
+    "MIRACLRetrieval": ["train"],
+    "CodeSearchNet": ["train"],
+}
+
+
+def q3e_instruct_loader(model_name_or_path, **kwargs):
+    model = InstructSentenceTransformerWrapper(
+        model_name_or_path,
+        revision=kwargs.pop("revision", None),
+        instruction_template=instruction_template,
+        apply_instruction_to_passages=False,
+        **kwargs,
+    )
+    encoder = model.model._first_module()
+    if encoder.auto_model.config._attn_implementation == "flash_attention_2":
+        # The Qwen3 code only use left padding in flash_attention_2 mode.
+        encoder.tokenizer.padding_side = "left"
+    return model
+
+
+Qwen3_Embedding_0B6 = ModelMeta(
+    loader=partial(  # type: ignore
+        q3e_instruct_loader,
+        model_name_or_path=os.environ.get("Q3E_0B6_PATH", "Qwen/Qwen3-Embedding-0.6B"),
+    ),
+    name="Qwen/Qwen3-Embedding-0.6B",
+    languages=multilingual_langs,
+    open_weights=True,
+    revision="b22da495047858cce924d27d76261e96be6febc0",  # Commit of @tomaarsen
+    release_date="2025-06-05",
+    n_parameters=595776512,
+    memory_usage_mb=2272,
+    embed_dim=1024,
+    max_tokens=32768,
+    license="apache-2.0",
+    reference="https://huggingface.co/Qwen/Qwen3-Embedding-0.6B",
+    similarity_fn_name="cosine",
+    framework=["Sentence Transformers", "PyTorch"],
+    use_instructions=True,
+    public_training_code=None,
+    public_training_data=None,
+    training_datasets=training_data,
+)
+
+Qwen3_Embedding_4B = ModelMeta(
+    loader=partial(  # type: ignore
+        q3e_instruct_loader,
+        model_name_or_path=os.environ.get("Q3E_4B_PATH", "Qwen/Qwen3-Embedding-4B"),
+    ),
+    name="Qwen/Qwen3-Embedding-4B",
+    languages=multilingual_langs,
+    open_weights=True,
+    revision="636cd9bf47d976946cdbb2b0c3ca0cb2f8eea5ff",  # Commit of @tomaarsen
+    release_date="2025-06-05",
+    n_parameters=4021774336,
+    memory_usage_mb=15341,
+    embed_dim=2560,
+    max_tokens=32768,
+    license="apache-2.0",
+    reference="https://huggingface.co/Qwen/Qwen3-Embedding-4B",
+    similarity_fn_name="cosine",
+    framework=["Sentence Transformers", "PyTorch"],
+    use_instructions=True,
+    public_training_code=None,
+    public_training_data=None,
+    training_datasets=training_data,
+)
+
+Qwen3_Embedding_8B = ModelMeta(
+    loader=partial(  # type: ignore
+        q3e_instruct_loader,
+        model_name_or_path=os.environ.get("Q3E_8B_PATH", "Qwen/Qwen3-Embedding-8B"),
+    ),
+    name="Qwen/Qwen3-Embedding-8B",
+    languages=multilingual_langs,
+    open_weights=True,
+    revision="4e423935c619ae4df87b646a3ce949610c66241c",  # Commit of @tomaarsen
+    release_date="2025-06-05",
+    n_parameters=7567295488,
+    memory_usage_mb=28866,
+    embed_dim=4096,
+    max_tokens=32768,
+    license="apache-2.0",
+    reference="https://huggingface.co/Qwen/Qwen3-Embedding-8B",
+    similarity_fn_name="cosine",
+    framework=["Sentence Transformers", "PyTorch"],
+    use_instructions=True,
+    public_training_code=None,
+    public_training_data=None,
+    training_datasets=training_data,
+)