gte-large-zh

General Text Embeddings (GTE) model. Towards General Text Embeddings with Multi-stage Contrastive Learning

The GTE models are trained by Alibaba DAMO Academy. They are mainly based on the BERT framework and currently offer different sizes of models for both Chinese and English Languages. The GTE models are trained on a large-scale corpus of relevance text pairs, covering a wide range of domains and scenarios. This enables the GTE models to be applied to various downstream tasks of text embeddings, including information retrieval, semantic textual similarity, text reranking, etc.

Model List

Models	Language	Max Sequence Length	Dimension	Model Size
GTE-large-zh	Chinese	512	1024	0.67GB
GTE-base-zh	Chinese	512	512	0.21GB
GTE-small-zh	Chinese	512	512	0.10GB
GTE-large	English	512	1024	0.67GB
GTE-base	English	512	512	0.21GB
GTE-small	English	512	384	0.10GB

Metrics

We compared the performance of the GTE models with other popular text embedding models on the MTEB (CMTEB for Chinese language) benchmark. For more detailed comparison results, please refer to the MTEB leaderboard.

Evaluation results on CMTEB

Model	Model Size (GB)	Embedding Dimensions	Sequence Length	Average (35 datasets)	Classification (9 datasets)	Clustering (4 datasets)	Pair Classification (2 datasets)	Reranking (4 datasets)	Retrieval (8 datasets)	STS (8 datasets)
gte-large-zh	0.65	1024	512	66.72	71.34	53.07	81.14	67.42	72.49	57.82
gte-base-zh	0.20	768	512	65.92	71.26	53.86	80.44	67.00	71.71	55.96
stella-large-zh-v2	0.65	1024	1024	65.13	69.05	49.16	82.68	66.41	70.14	58.66
stella-large-zh	0.65	1024	1024	64.54	67.62	48.65	78.72	65.98	71.02	58.3
bge-large-zh-v1.5	1.3	1024	512	64.53	69.13	48.99	81.6	65.84	70.46	56.25
stella-base-zh-v2	0.21	768	1024	64.36	68.29	49.4	79.96	66.1	70.08	56.92
stella-base-zh	0.21	768	1024	64.16	67.77	48.7	76.09	66.95	71.07	56.54
piccolo-large-zh	0.65	1024	512	64.11	67.03	47.04	78.38	65.98	70.93	58.02
piccolo-base-zh	0.2	768	512	63.66	66.98	47.12	76.61	66.68	71.2	55.9
gte-small-zh	0.1	512	512	60.04	64.35	48.95	69.99	66.21	65.50	49.72
bge-small-zh-v1.5	0.1	512	512	57.82	63.96	44.18	70.4	60.92	61.77	49.1
m3e-base	0.41	768	512	57.79	67.52	47.68	63.99	59.54	56.91	50.47
text-embedding-ada-002(openai)	-	1536	8192	53.02	64.31	45.68	69.56	54.28	52.0	43.35

Usage

Code example

import torch.nn.functional as F
from torch import Tensor
from transformers import AutoTokenizer, AutoModel

input_texts = [
    "中国的首都是哪里",
    "你喜欢去哪里旅游",
    "北京",
    "今天中午吃什么"
]

tokenizer = AutoTokenizer.from_pretrained("thenlper/gte-large-zh")
model = AutoModel.from_pretrained("thenlper/gte-large-zh")

# Tokenize the input texts
batch_dict = tokenizer(input_texts, max_length=512, padding=True, truncation=True, return_tensors='pt')

outputs = model(**batch_dict)
embeddings = outputs.last_hidden_state[:, 0]
 
# (Optionally) normalize embeddings
embeddings = F.normalize(embeddings, p=2, dim=1)
scores = (embeddings[:1] @ embeddings[1:].T) * 100
print(scores.tolist())

Use with sentence-transformers:

from sentence_transformers import SentenceTransformer
from sentence_transformers.util import cos_sim

sentences = ['That is a happy person', 'That is a very happy person']

model = SentenceTransformer('thenlper/gte-large-zh')
embeddings = model.encode(sentences)
print(cos_sim(embeddings[0], embeddings[1]))

Limitation

This model exclusively caters to Chinese texts, and any lengthy texts will be truncated to a maximum of 512 tokens.

Citation

If you find our paper or models helpful, please consider citing them as follows:

@article{li2023towards,
  title={Towards general text embeddings with multi-stage contrastive learning},
  author={Li, Zehan and Zhang, Xin and Zhang, Yanzhao and Long, Dingkun and Xie, Pengjun and Zhang, Meishan},
  journal={arXiv preprint arXiv:2308.03281},
  year={2023}
}