openbmb/MiniCPM-Reranker-Light

MiniCPM-Reranker-Light

MiniCPM-Reranker-Light 是面壁智能与清华大学自然语言处理实验室（THUNLP）、东北大学信息检索小组（NEUIR）共同开发的中英双语言文本重排序模型，有如下特点：

• 出色的中文、英文重排序能力。
• 出色的中英跨语言重排序能力。
• 支持长文本（最长8192token）。

MiniCPM-Reranker-Light 基于 MiniCPM-1B-sft-bf16 训练，结构上采取双向注意力。采取多阶段训练方式，共使用包括开源数据、机造数据、闭源数据在内的约 500 万条训练数据。

欢迎关注 UltraRAG 系列：

• 检索模型：MiniCPM-Embedding-Light
• 重排模型：MiniCPM-Reranker-Light
• 领域自适应RAG框架：UltraRAG

MiniCPM-Reranker-Light is a bilingual & cross-lingual text re-ranking model developed by ModelBest Inc. , THUNLP and NEUIR , featuring:

• Exceptional Chinese and English re-ranking capabilities.
• Outstanding cross-lingual re-ranking capabilities between Chinese and English.
• Long-text support (up to 8192 tokens).

MiniCPM-Reranker-Light is trained based on MiniCPM-1B-sft-bf16 and incorporates bidirectional attention in its architecture. The model underwent multi-stage training using approximately 6 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

模型信息 Model Information

• 模型大小：1.2B

• 最大输入token数：8192

• Model Size: 1.2B

• Max Input Tokens: 8192

使用方法 Usage

输入格式 Input Format

本模型支持指令，输入格式如下：

MiniCPM-Reranker-Light supports instructions in the following format:

<s>Instruction: {{ instruction }} Query: {{ query }}</s>{{ document }}

例如：

For example:

<s>Instruction: 为这个医学问题检索相关回答。Query: 咽喉癌的成因是什么？</s>（文档省略）

<s>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.</s>(document omitted)

也可以不提供指令，即采取如下格式：

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

<s>Query: {{ query }}</s>{{ document }}

我们在BEIR与C-MTEB/Retrieval上测试时使用的指令见 instructions.json，其他测试不使用指令。

When running evaluation on BEIR and C-MTEB/Retrieval, we use instructions in instructions.json. For other evaluations, we do not use instructions.

环境要求 Requirements

transformers==4.37.2

示例脚本 Demo

Huggingface Transformers

from transformers import AutoModelForSequenceClassification
import torch

model_name = "OpenBMB/MiniCPM-Reranker-Light"
model = AutoModelForSequenceClassification.from_pretrained(model_name, trust_remote_code=True, torch_dtype=torch.float16).to("cuda")
# You can also use the following code to use flash_attention_2
# model = AutoModelForSequenceClassification.from_pretrained(model_name, trust_remote_code=True,attn_implementation="flash_attention_2", torch_dtype=torch.float16).to("cuda")
model.eval()

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

rerank_score = model.rerank(query, passages,query_instruction="Query:", batch_size=32, max_length=1024)
print(rerank_score) #[0.01791382 0.00024533]


sentence_pairs = [[f"Query: {query}", doc] for doc in passages]
scores = model.compute_score(sentence_pairs, batch_size=32, max_length=1024)
print(scores) #[0.01791382 0.00024533]

Sentence Transformer

from sentence_transformers import CrossEncoder
from transformers import LlamaTokenizer
import torch

model_name = "OpenBMB/MiniCPM-Reranker-Light"
model = CrossEncoder(model_name,max_length=1024,trust_remote_code=True, automodel_args={"torch_dtype": torch.float16})
# You can also use the following code to use flash_attention_2
#model = CrossEncoder(model_name,max_length=1024,trust_remote_code=True, automodel_args={"attn_implementation":"flash_attention_2","torch_dtype": torch.float16})
model.tokenizer.padding_side = "right"

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

INSTRUCTION = "Query: "
query = INSTRUCTION + query

sentence_pairs = [[query, doc] for doc in passages]

scores = model.predict(sentence_pairs, convert_to_tensor=True).tolist()
rankings = model.rank(query, passages, return_documents=True, convert_to_tensor=True)

print(scores) # [0.017913818359375, 0.0002453327178955078]
for ranking in rankings:
    print(f"Score: {ranking['score']:.4f}, Corpus: {ranking['text']}")
  
# Score: 0.0179, Corpus: beijing
# Score: 0.0002, Corpus: shanghai

Infinity

import asyncio
from infinity_emb import AsyncEngineArray, EngineArgs, AsyncEmbeddingEngine
query = "中国的首都是哪里？" # "What is the capital of China?"
docs = ["beijing", "shanghai"] # "北京", "上海"

INSTRUCTION = "Query:"
query = f"{INSTRUCTION} {query}"

array = AsyncEngineArray.from_args(
  [EngineArgs(model_name_or_path = "OpenBMB/MiniCPM-Reranker-Light", engine="torch", dtype="float16", bettertransformer=False, trust_remote_code=True, model_warmup=False)]
)

async def rerank(engine: AsyncEmbeddingEngine): 
    async with engine:
        ranking, usage = await engine.rerank(query=query, docs=docs)
        print(list(zip(ranking, docs)))

asyncio.run(rerank(array[0])) # [(RerankReturnType(relevance_score=0.017917344, document='beijing', index=0), 'beijing'), (RerankReturnType(relevance_score=0.00024729347, document='shanghai', index=1), 'shanghai')]

FlagEmbedding

from FlagEmbedding import FlagReranker
model_name = "OpenBMB/MiniCPM-Reranker-Light"
model = FlagReranker(model_name, use_fp16=True, query_instruction_for_rerank="Query: ", trust_remote_code=True)
# You can hack the __init__() method of the FlagEmbedding BaseReranker class to use flash_attention_2 for faster inference
#  self.model = AutoModelForSequenceClassification.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
#         )
model.tokenizer.padding_side = "right"

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

sentence_pairs = [[query, doc] for doc in passages]

scores = model.compute_score(sentence_pairs,normalize=True)
print(scores) # [0.01791734476747132, 0.0002472934613244585]

实验结果 Evaluation Results

中文与英文重排序结果 CN/EN Re-ranking Results

中文对bge-large-zh-v1.5检索的top-100进行重排，英文对bge-large-en-v1.5检索的top-100进行重排。

We re-rank top-100 docments from bge-large-zh-v1.5 in C-MTEB/Retrieval and from bge-large-en-v1.5 in BEIR.

模型 Model	C-MTEB/Retrieval (NDCG@10)	BEIR (NDCG@10)
bge-large-zh-v1.5（Retriever for Chinese）	70.46	-
bge-large-en-v1.5（Retriever for English）	-	54.29
bge-reranker-v2-m3	71.82	55.36
bge-reranker-v2-minicpm-28	73.51	59.86
bge-reranker-v2-gemma	71.74	60.71
bge-reranker-v2.5-gemma2	-	63.67
MiniCPM-Reranker	76.79	61.32
MiniCPM-Reranker-Light	76.19	61.34

中英跨语言重排序结果 CN-EN Cross-lingual Re-ranking Results

对bge-m3(Dense)检索的top100进行重排。

We re-rank top-100 documents from bge-m3 (Dense).

模型 Model	MKQA En-Zh_CN (Recall@20)	NeuCLIR22 (NDCG@10)	NeuCLIR23 (NDCG@10)
bge-m3 (Dense)（Retriever）	66.4	30.49	41.09
jina-reranker-v2-base-multilingual	69.33	36.66	50.03
bge-reranker-v2-m3	69.75	40.98	49.67
gte-multilingual-reranker-base	68.51	38.74	45.3
MiniCPM-Reranker	71.73	43.65	50.59
MiniCPM-Reranker-Light	71.34	46.04	51.86

许可证 License

• 本仓库中代码依照 Apache-2.0 协议开源。
• MiniCPM-Reranker-Light 模型权重的使用则需要遵循 MiniCPM 模型协议。
• MiniCPM-Reranker-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-Reranker-Light model weights must strictly follow MiniCPM Model License.md.
• The models and weights of MiniCPM-Reranker-Light are completely free for academic research. After filling out a "questionnaire" for registration, MiniCPM-Reranker-Light weights are also available for free commercial use.