Chuxin-Embedding

Chuxin-Embedding 是专为增强中文文本检索能力而设计的嵌入模型。它基于bge-m3-retromae[1]，实现了预训练、微调、精调全流程。该模型在来自各个领域的大量语料库上进行训练，语料库的批量非常大。截至 2024 年 9 月 14 日, Chuxin-Embedding 在检索任务中表现出色，在 C-MTEB 中文检索排行榜上排名第一，领先的性能得分为 77.88，在AIR-Bench中文检索+重排序公开排行榜上排名第一，领先的性能得分为 64.78。

Chuxin-Embedding is a specially designed embedding model aimed at enhancing the capability of Chinese text retrieval. It is based on bge-m3-retromae[1] and implements the entire process of pre-training, fine-tuning, and refining. This model has been trained on a vast amount of corpora from various fields. As of September 14, 2024, Chuxin-Embedding has shown outstanding performance in retrieval tasks. It ranks first on the C-MTEB Chinese Retrieval Leaderboard with a leading performance score of 77.88 and also ranks first on the AIR-Bench Chinese Retrieval + Re-ranking Public Leaderboard with a leading performance score of 64.78.

News

• 2024/10/18: LLM生成及数据清洗 Code 。

• 2024/9/14: 团队的RAG框架欢迎试用 ragnify 。

• 2024/9/14: LLM generation and data clean Code .

• 2024/9/14: The team's RAG framework is available for trial ragnify .

Training Details

基于bge-m3-retromae[1]，主要改动如下：

• 基于bge-m3-retromae[1]在亿级数据上预训练。
  • 使用BGE pretrain Code 完成预训练。
• 在收集的公开亿级检索数据集上实现了微调。
  • 使用BGE finetune Code 完成微调。
• 在收集的公开百万级检索数据集和百万级LLM合成数据集上实现了精调。
  • 使用BGE finetune Code 和 BGE unified_finetune Code 完成精调。
  • 通过 LLM (QWEN-72B) 进行数据生成，使用 LLM 为message生成新query
  • 数据清洗：
    • 简单的基于规则清洗
    • LLM判断是否可作为搜索引擎查询的query
    • rerank模型对(query,message)评分，舍弃pos中的负例，neg中的正例

Based on bge-m3-retromae[1], the main modifications are as follows:

• Pre-trained on a billion-level dataset based on bge-m3-retromae[1].
  • Pre-training is completed using BGE pretrain Code .
• Fine-tuned on a publicly collected billion-level retrieval dataset.
  • Fine-tuning is completed using BGE finetune Code.
• Refined on a publicly collected million-level retrieval dataset and a million-level LLM synthetic dataset.
  • Refining is completed using BGE finetune Code and BGE unified_finetune Code.
  • Data generation is performed through LLM (QWEN-72B), using LLM to generate new query for messages.
  • Data cleaning:
    • Simple rule-based cleaning
    • LLM to determine whether a query can be used as a search engine query
    • The rerank model scores (query, message) pairs, discarding negative examples in the positive set and positive examples in the negative set.

Collect more data for retrieval-type tasks

1. 预训练数据
   • ChineseWebText、 oasis、 oscar、 SkyPile、 wudao
2. 微调数据
   • MTP 、webqa、nlpcc、csl、bq、atec、ccks
3. 精调数据
   • BGE-M3 、Huatuo26M-Lite 、covid ...
   • LLM 合成（BGE-M3 、Huatuo26M-Lite 、covid、wudao、wanjuan_news、mnbvc_news_wiki、mldr、medical QA...）

Performance

C_MTEB RETRIEVAL

AIR-Bench

Generate Embedding for text

#pip install -U FlagEmbedding

from FlagEmbedding import FlagModel

model = FlagModel('chuxin-llm/Chuxin-Embedding',
                   query_instruction_for_retrieval="为这个句子生成表示以用于检索相关文章：",
                   use_fp16=True)

sentences_1 = ["样例数据-1", "样例数据-2"]
sentences_2 = ["样例数据-3", "样例数据-1"]

embeddings_1 = model.encode(sentences_1)
embeddings_2 = model.encode(sentences_2)
similarity = embeddings_1 @ embeddings_2.T
print(similarity)

Reference

1. https://huggingface.co/BAAI/bge-m3-retromae
2. https://github.com/FlagOpen/FlagEmbedding/tree/master/FlagEmbedding/BGE_M3
3. https://github.com/FlagOpen/FlagEmbedding/tree/master/FlagEmbedding/baai_general_embedding