MiniCPM-Embedding-Light incorporates bidirectional attention and Weighted Mean Pooling [1] in its architecture. The model underwent multi-stage training using approximately 260 million training examples, including open-source, synthetic, and proprietary data.

We also invite you to explore the UltraRAG series:

Retrieval Model: MiniCPM-Embedding-Light
Re-ranking Model: MiniCPM-Reranker-Light
Domain Adaptive RAG Framework: UltraRAG

[1] Muennighoff, N. (2022). Sgpt: Gpt sentence embeddings for semantic search. arXiv preprint arXiv:2202.08904. [2] Kusupati, Aditya, et al. "Matryoshka representation learning." Advances in Neural Information Processing Systems 35 (2022): 30233-30249.

模型信息 Model Information

模型大小：440M
嵌入维度：1024
最大输入token数：8192
Model Size: 440M
Embedding Dimension: 1024
Max Input Tokens: 8192

使用方法 Usage

输入格式 Input Format

本模型支持 query 侧指令，格式如下：

MiniCPM-Embedding-Light supports query-side instructions in the following format:

Instruction: {{ instruction }} Query: {{ query }}

例如：

For example:

Instruction: 为这个医学问题检索相关回答。Query: 咽喉癌的成因是什么？

Instruction: Given a claim about climate change, retrieve documents that support or refute the claim. Query: However the warming trend is slower than most climate models have forecast.

也可以不提供指令，即采取如下格式：

MiniCPM-Embedding-Light also works in instruction-free mode in the following format:

Query: {{ query }}

环境要求 Requirements

transformers==4.37.2

示例脚本 Demo

Huggingface Transformers

python

from transformers import AutoModel
import torch

model_name = "openbmb/MiniCPM-Embedding-Light"
model = AutoModel.from_pretrained(model_name, trust_remote_code=True, torch_dtype=torch.float16).to("cuda")

# you can use flash_attention_2 for faster inference
# model = AutoModel.from_pretrained(model_name, trust_remote_code=True, attn_implementation="flash_attention_2", torch_dtype=torch.float16).to("cuda") 

model.eval()

queries = ["MiniCPM-o 2.6 A GPT-4o Level MLLM for Vision, Speech and Multimodal Live Streaming on Your Phone"]
passages = ["MiniCPM-o 2.6 is the latest and most capable model in the MiniCPM-o series. The model is built in an end-to-end fashion based on SigLip-400M, Whisper-medium-300M, ChatTTS-200M, and Qwen2.5-7B with a total of 8B parameters. It exhibits a significant performance improvement over MiniCPM-V 2.6, and introduces new features for real-time speech conversation and multimodal live streaming."]

embeddings_query_dense, embeddings_query_sparse = model.encode_query(queries, return_sparse_vectors=True)
embeddings_doc_dense, embeddings_doc_sparse = model.encode_corpus(passages, return_sparse_vectors=True)

dense_scores = (embeddings_query_dense @ embeddings_doc_dense.T)
print(dense_scores.tolist())  # [[0.6512398719787598]]
print(model.compute_sparse_score_dicts(embeddings_query_sparse,  embeddings_doc_sparse)) # [[0.27202296]]

dense_scores, sparse_scores, mixed_scores = model.compute_score(queries, passages)
print(dense_scores) # [[0.65123993]]
print(sparse_scores) # [[0.27202296]]
print(mixed_scores) # [[0.73284686]]

Sentence Transformers

python

import torch
from sentence_transformers import SentenceTransformer


model_name = "openbmb/MiniCPM-Embedding-Light"
model = SentenceTransformer(model_name, trust_remote_code=True, model_kwargs={"torch_dtype": torch.float16})

# you can use flash_attention_2 for faster inference
# model = SentenceTransformer(model_name, trust_remote_code=True, model_kwargs={"attn_implementation": "flash_attention_2", "torch_dtype": torch.float16})

queries = ["中国的首都是哪里？"] # "What is the capital of China?"
passages = ["beijing", "shanghai"] # "北京", "上海"

INSTRUCTION = "Query: "

embeddings_query = model.encode(queries, prompt=INSTRUCTION)
embeddings_doc = model.encode(passages)

scores = (embeddings_query @ embeddings_doc.T)
print(scores.tolist())  # [[0.40356746315956116, 0.36183440685272217]]

Infinity

python

import asyncio
from infinity_emb import AsyncEngineArray, EngineArgs, AsyncEmbeddingEngine
import numpy as np

array = AsyncEngineArray.from_args([
  EngineArgs(model_name_or_path = "openbmb/MiniCPM-Embedding-Light", engine="torch", dtype="float16", bettertransformer=False, pooling_method="mean", trust_remote_code=True),
])
queries = ["中国的首都是哪里？"] # "What is the capital of China?"
passages = ["beijing", "shanghai"] # "北京", "上海"

INSTRUCTION = "Query:"
queries = [f"{INSTRUCTION} {query}" for query in queries]


async def embed_text(engine: AsyncEmbeddingEngine,sentences): 
    async with engine: 
        embeddings, usage = await engine.embed(sentences=sentences)
    return embeddings

queries_embedding = asyncio.run(embed_text(array[0],queries))
passages_embedding = asyncio.run(embed_text(array[0],passages))

scores = (np.array(queries_embedding) @ np.array(passages_embedding).T)
print(scores.tolist())  # [[0.40356746315956116, 0.36183443665504456]]

FlagEmbedding

python

from FlagEmbedding import FlagModel


model = FlagModel("openbmb/MiniCPM-Embedding-Light", 
                          query_instruction_for_retrieval="Query: ",
                          pooling_method="mean",
                          trust_remote_code=True,
                          normalize_embeddings=True,
                          use_fp16=True)
# You can hack the __init__() method of the FlagEmbedding BaseEmbedder class to use flash_attention_2 for faster inference
#  self.model = AutoModel.from_pretrained(
#             model_name_or_path,
#             trust_remote_code=trust_remote_code,
#             cache_dir=cache_dir,
#             # torch_dtype=torch.float16, # we need to add this line to use fp16
#             # attn_implementation="flash_attention_2", # we need to add this line to use flash_attention_2
#         )

queries = ["中国的首都是哪里？"] # "What is the capital of China?"
passages = ["beijing", "shanghai"] # "北京", "上海"


embeddings_query = model.encode_queries(queries)
embeddings_doc = model.encode_corpus(passages)

scores = (embeddings_query @ embeddings_doc.T)
print(scores.tolist())  # [[0.40356746315956116, 0.36183440685272217]]

实验结果 Evaluation Results

中文与英文检索结果 CN/EN Retrieval Results

模型 Model	C-MTEB/Retrieval(NDCG@10)	BEIR(NDCG@10)
bge-large-zh-v1.5	70.46	-
gte-large-zh	72.49	-
Conan-embedding-v1	76.67
bge-large-en-v1.5	-	54.29
modernbert-embed-large	-	54.36
snowflake-arctic-embed-l	-	55.98
gte-en-large-v1.5	-	57.91
me5-large	63.66	51.43
bge-m3(Dense)	65.43	48.82
gte-multilingual-base(Dense)	71.95	51.08
jina-embeddings-v3	68.60	53.88
gte-Qwen2-1.5B-instruct	71.86	58.29
MiniCPM-Embedding	76.76	58.56
MiniCPM-Embedding-Light(Dense)	72.71	55.27
MiniCPM-Embedding-Light(Dense+Sparse)	73.13	56.31
MiniCPM-Embedding-Light(Dense+Sparse)+MiniCPM-Reranker-Light	76.34	61.49

中英跨语言检索结果 CN-EN Cross-lingual Retrieval Results

模型 Model	MKQA En-Zh_CN (Recall@20)	NeuCLIR22 (NDCG@10)	NeuCLIR23 (NDCG@10)
me5-large	44.3	9.01	25.33
bge-m3(Dense)	66.4	30.49	41.09
gte-multilingual-base(Dense)	68.2	39.46	45.86
MiniCPM-Embedding	72.95	52.65	49.95
MiniCPM-Embedding-Light(Dense)	68.29	41.17	45.83
MiniCPM-Embedding-Light(Dense)+MiniCPM-Reranker-Light	71.86	54.32	56.50

许可证 License

本仓库中代码依照 Apache-2.0 协议开源。
MiniCPM-Embedding-Light 模型权重的使用则需要遵循 MiniCPM 模型协议。
MiniCPM-Embedding-Light 模型权重对学术研究完全开放。如需将模型用于商业用途，请填写此问卷。

The code in this repo is released under the Apache-2.0 License.
The usage of MiniCPM-Embedding-Light model weights must strictly follow MiniCPM Model License.md.
The models and weights of MiniCPM-Embedding-Light are completely free for academic research. After filling out a "questionnaire" for registration, MiniCPM-Embedding-Light weights are also available for free commercial use.