fix: Added Misc Chinese models (#1819)

* Added moka and piccolo models to overview file

* Added Text2Vec models

* Added various Chinese embedding models

---------

Co-authored-by: Isaac Chung <chungisaac1217@gmail.com>

Author: x-tabdeveloping

mteb/models/misc_models.py

@@ -1738,3 +1738,91 @@
     training_datasets=None,  # They don't specify
     superseded_by=None,
 )
+xiaobu_embedding = ModelMeta(
+    name="lier007/xiaobu-embedding",
+    revision="59c79d82eb5223cd9895f6eb8e825c7fa10e4e92",
+    release_date="2024-01-09",
+    languages=["zho_Hans"],
+    loader=None,
+    n_parameters=326 * 1e6,
+    memory_usage=None,
+    max_tokens=512,
+    embed_dim=1024,
+    license="not specified",
+    open_weights=True,
+    public_training_data=False,
+    public_training_code=None,
+    framework=["PyTorch", "Sentence Transformers"],
+    reference="https://huggingface.co/lier007/xiaobu-embedding",
+    similarity_fn_name="cosine",
+    use_instructions=None,
+    training_datasets=None,  # Finetuned from GTE, none of them disclose training data
+    superseded_by=None,
+    adapted_from="thenlper/gte-large-zh",
+)
+xiaobu_embedding_v2 = ModelMeta(
+    name="lier007/xiaobu-embedding-v2",
+    revision="1912f2e59a5c2ef802a471d735a38702a5c9485e",
+    release_date="2024-06-30",
+    languages=["zho_Hans"],
+    loader=None,
+    n_parameters=326 * 1e6,
+    memory_usage=None,
+    max_tokens=512,
+    embed_dim=768,
+    license="not specified",
+    open_weights=True,
+    public_training_data=False,
+    public_training_code=None,
+    framework=["PyTorch", "Sentence Transformers"],
+    reference="https://huggingface.co/lier007/xiaobu-embedding-v2",
+    similarity_fn_name="cosine",
+    use_instructions=None,
+    training_datasets=None,  # Finetuned from piccolo-embedding, none of them say
+    superseded_by=None,
+    adapted_from="sensenova/piccolo-base-zh",
+)
+yinka_embedding = ModelMeta(
+    name="Classical/Yinka",
+    revision="59c79d82eb5223cd9895f6eb8e825c7fa10e4e92",
+    release_date="2024-01-09",
+    languages=["zho_Hans"],
+    loader=None,
+    n_parameters=326 * 1e6,
+    memory_usage=None,
+    max_tokens=512,
+    embed_dim=1024,
+    license="not specified",
+    open_weights=True,
+    public_training_data=False,
+    public_training_code=None,
+    framework=["PyTorch", "Sentence Transformers"],
+    reference="https://huggingface.co/Classical/Yinka",
+    similarity_fn_name="cosine",
+    use_instructions=None,
+    training_datasets=None,  # Not disclosed
+    superseded_by=None,
+    adapted_from="dunzhang/stella-mrl-large-zh-v3.5-1792d",
+)
+conan_embedding = ModelMeta(
+    name="TencentBAC/Conan-embedding-v1",
+    revision="bb9749a57d4f02fd71722386f8d0f5a9398d7eeb",
+    release_date="2024-08-22",
+    languages=["zho_Hans"],
+    loader=None,
+    n_parameters=326 * 1e6,
+    memory_usage=None,
+    max_tokens=512,
+    embed_dim=768,
+    license="cc-by-nc-4.0",
+    open_weights=True,
+    public_training_data=False,
+    public_training_code=None,
+    framework=["PyTorch", "Sentence Transformers"],
+    reference="https://huggingface.co/Classical/Yinka",
+    similarity_fn_name="cosine",
+    use_instructions=None,
+    # source: https://arxiv.org/pdf/2408.15710
+    training_datasets=None,  # They "scraped" things from the internet, we don't know, could be leakage
+    superseded_by=None,
+)

mteb/models/overview.py

@@ -36,6 +36,7 @@
     nomic_models,
     nvidia_models,
     openai_models,
+    piccolo_models,
     promptriever_models,
     repllama_models,
     rerankers_custom,
@@ -44,6 +45,7 @@
     salesforce_models,
     sentence_transformers_models,
     stella_models,
+    text2vec_models,
     uae_models,
     voyage_models,
 )
@@ -69,11 +71,13 @@
     llm2vec_models,
     mxbai_models,
     model2vec_models,
+    moka_models,
     misc_models,
     nomic_models,
     no_instruct_sentence_models,
     nvidia_models,
     openai_models,
+    piccolo_models,
     promptriever_models,
     repllama_models,
     rerankers_custom,
@@ -88,6 +92,7 @@
     jina_models,
     jasper_models,
     uae_models,
+    text2vec_models,
     stella_models,
     uae_models,
     voyage_models,

mteb/models/text2vec_models.py

@@ -0,0 +1,103 @@
+"""Implementation of Text2Vec models"""
+
+from __future__ import annotations
+
+from mteb.model_meta import ModelMeta
+
+# I couldn't find the large model on HF for some reason
+text2vec_base_chinese = ModelMeta(
+    name="shibing624/text2vec-base-chinese",
+    languages=["zho-Hans"],
+    open_weights=True,
+    revision="183bb99aa7af74355fb58d16edf8c13ae7c5433e",
+    release_date="2022-01-23",
+    n_parameters=102 * 1e6,
+    memory_usage=None,
+    embed_dim=768,
+    license="apache-2.0",
+    max_tokens=512,
+    reference="https://huggingface.co/shibing624/text2vec-base-chinese",
+    similarity_fn_name="cosine",
+    framework=["Sentence Transformers", "PyTorch"],
+    use_instructions=False,
+    superseded_by=None,
+    adapted_from=None,
+    public_training_code=False,  # Couldn't find it
+    public_training_data=True,
+    training_datasets={
+        # source: https://huggingface.co/shibing624/text2vec-base-chinese
+        # Not in MTEB
+        # - shibing624/nli-zh-all/text2vec-base-chinese-sentence-dataset
+        # (Could have overlaps I'm not aware of)
+    },
+)
+
+text2vec_base_chinese_paraphrase = ModelMeta(
+    name="shibing624/text2vec-base-chinese-paraphrase",
+    languages=["zho-Hans"],
+    open_weights=True,
+    revision="e90c150a9c7fb55a67712a766d6820c55fb83cdd",
+    release_date="2023-06-19",
+    n_parameters=118 * 1e6,
+    memory_usage=None,
+    embed_dim=768,
+    license="apache-2.0",
+    max_tokens=512,
+    reference="https://huggingface.co/shibing624/text2vec-base-chinese-paraphrase",
+    similarity_fn_name="cosine",
+    framework=["Sentence Transformers", "PyTorch"],
+    use_instructions=False,
+    superseded_by=None,
+    adapted_from=None,
+    public_training_code=False,  # Couldn't find it
+    public_training_data=True,
+    training_datasets={
+        # source: https://huggingface.co/shibing624/text2vec-base-chinese
+        # Not in MTEB
+        # - shibing624/nli-zh-all/text2vec-base-chinese-paraphrase
+        # (Could have overlaps I'm not aware of)
+    },
+)
+
+
+text2vec_multi_langs = [
+    "deu-Latn",  # German (de)
+    "eng-Latn",  # English (en)
+    "spa-Latn",  # Spanish (es)
+    "fra-Latn",  # French (fr)
+    "ita-Latn",  # Italian (it)
+    "nld-Latn",  # Dutch (nl)
+    "pol-Latn",  # Polish (pl)
+    "por-Latn",  # Portuguese (pt)
+    "rus-Cyrl",  # Russian (ru)
+    "zho-Hans",  # Chinese (Simplified, zh)
+]
+text2vec_base_multilingual = ModelMeta(
+    name="shibing624/text2vec-base-multilingual",
+    languages=text2vec_multi_langs,
+    open_weights=True,
+    revision="6633dc49e554de7105458f8f2e96445c6598e9d1",
+    release_date="2023-06-22",
+    # While it can be loaded with SBERT, it has one suspicious file according to huggingface
+    # So probably best not to.
+    loader=None,
+    n_parameters=118 * 1e6,
+    memory_usage=None,
+    embed_dim=384,
+    license="apache-2.0",
+    max_tokens=256,
+    reference="https://huggingface.co/shibing624/text2vec-base-chinese-paraphrase",
+    similarity_fn_name="cosine",
+    framework=["Sentence Transformers", "PyTorch"],
+    use_instructions=False,
+    superseded_by=None,
+    adapted_from="sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2",
+    public_training_code=False,  # Couldn't find it
+    public_training_data=True,
+    training_datasets={
+        # source: https://huggingface.co/shibing624/text2vec-base-chinese
+        # Not in MTEB
+        # - shibing624/nli-zh-all/tree/main/text2vec-base-multilingual-dataset
+        # # (Could have overlaps I'm not aware of)
+    },
+)