Language & LLMs
New Chinese text-embedding models beat prior Chinese models by more than 10 percent on C-MTEB, a 35-dataset benchmark

In a 2024 paper, researchers from the Beijing Academy of Artificial Intelligence, Renmin University of China and Université de Montréal introduced C-Pack, a set of resources for general Chinese text embeddings. A text embedding is a numerical representation of words or sentences that lets a computer compare meaning. The paper reports that its BGE models outperformed all prior Chinese text embeddings on C-MTEB, the benchmark included in the release, by more than 10 percent at the time of release.
What is in the C-Pack release?
C-Pack has three main parts. C-MTP is a large training dataset that combines vast unlabeled corpora with high-quality labeled corpora. C-MTEB is a Chinese text-embedding benchmark covering 6 tasks and 35 datasets. BGE is a family of embedding models in several sizes. The authors also released their data and models for English text embeddings.
What did the researchers report for Chinese and English performance?
For Chinese, the BGE models beat all earlier Chinese text embeddings on C-MTEB by more than 10 percent at the time of release. For English, the paper reports that its models achieved state-of-the-art performance on the MTEB benchmark. The released English data is 2 times larger than the Chinese data. The paper describes both Chinese and English datasets as the largest public release of training data for text embeddings.
FAQ
What is a text embedding in natural language processing?
A text embedding is a numerical representation of words or sentences that lets a computer compare meaning. In natural language processing, texts with similar meanings can be placed close together in this numeric form.
What does C-MTEB measure?
C-MTEB is the Chinese benchmark included in the release. It covers 6 tasks and 35 datasets for Chinese text embeddings.
What is BGE?
BGE is a family of embedding models in the C-Pack release. The models come in multiple sizes and were used for the reported Chinese and English results.
What natural language processing techniques does the paper describe?
The abstract does not name individual techniques. It reports that the authors integrated and optimized the entire training method suite for BGE.
Limits and what's next
The Chinese performance result comes from one paper and one benchmark released alongside the models, C-MTEB, which covers 6 tasks and 35 datasets. The more than 10 percent gain is relative to earlier Chinese text embeddings at the time of release, so later comparisons are outside the reported result. The abstract describes the training data as vast unlabeled corpora combined with high-quality labeled corpora, but does not explain the curation steps in detail. This is one study, and the abstract reports no independent replication.
Source: Shitao Xiao et al., “C-Pack: Packed Resources For General Chinese Embeddings,” International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR), 2024.