gte-large-en-v1.5

We introduce gte-v1.5 series, upgraded gte embeddings that support the context length of up to 8192, while further enhancing model performance. The models are built upon the transformer++ encoder backbone (BERT + RoPE + GLU).

The gte-v1.5 series achieve state-of-the-art scores on the MTEB benchmark within the same model size category and prodvide competitive on the LoCo long-context retrieval tests (refer to Evaluation).

We also present the gte-Qwen1.5-7B-instruct, a SOTA instruction-tuned multi-lingual embedding model that ranked 2nd in MTEB and 1st in C-MTEB.

• Developed by: Institute for Intelligent Computing, Alibaba Group
• Model type: Text Embeddings
• Paper: Coming soon.

Model list

Models	Language	Model Size	Max Seq. Length	Dimension	MTEB-en	LoCo
gte-Qwen1.5-7B-instruct	Multiple	7720	32768	4096	67.34	87.57
gte-large-en-v1.5	English	434	8192	1024	65.39	86.71
gte-base-en-v1.5	English	137	8192	768	64.11	87.44

How to Get Started with the Model

Use the code below to get started with the model.

# Requires transformers>=4.36.0

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

input_texts = [
    "what is the capital of China?",
    "how to implement quick sort in python?",
    "Beijing",
    "sorting algorithms"
]

model_path = 'Alibaba-NLP/gte-large-en-v1.5'
tokenizer = AutoTokenizer.from_pretrained(model_path)
model = AutoModel.from_pretrained(model_path, trust_remote_code=True)

# Tokenize the input texts
batch_dict = tokenizer(input_texts, max_length=8192, 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())

It is recommended to install xformers and enable unpadding for acceleration, refer to enable-unpadding-and-xformers.

Use with sentence-transformers:

# Requires sentence_transformers>=2.7.0

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('Alibaba-NLP/gte-large-en-v1.5', trust_remote_code=True)
embeddings = model.encode(sentences)
print(cos_sim(embeddings[0], embeddings[1]))

Use with transformers.js:

// npm i @xenova/transformers
import { pipeline, dot } from '@xenova/transformers';

// Create feature extraction pipeline
const extractor = await pipeline('feature-extraction', 'Alibaba-NLP/gte-large-en-v1.5', {
    quantized: false, // Comment out this line to use the quantized version
});

// Generate sentence embeddings
const sentences = [
    "what is the capital of China?",
    "how to implement quick sort in python?",
    "Beijing",
    "sorting algorithms"
]
const output = await extractor(sentences, { normalize: true, pooling: 'cls' });

// Compute similarity scores
const [source_embeddings, ...document_embeddings ] = output.tolist();
const similarities = document_embeddings.map(x => 100 * dot(source_embeddings, x));
console.log(similarities); // [41.86354093370361, 77.07076371259589, 37.02981979677899]

Training Details

Training Data

• Masked language modeling (MLM): c4-en
• Weak-supervised contrastive (WSC) pre-training: GTE pre-training data
• Supervised contrastive fine-tuning: GTE(https://arxiv.org/pdf/2308.03281.pdf) fine-tuning data

Training Procedure

To enable the backbone model to support a context length of 8192, we adopted a multi-stage training strategy. The model first undergoes preliminary MLM pre-training on shorter lengths. And then, we resample the data, reducing the proportion of short texts, and continue the MLM pre-training.

The entire training process is as follows:

• MLM-512: lr 2e-4, mlm_probability 0.3, batch_size 4096, num_steps 300000, rope_base 10000
• MLM-2048: lr 5e-5, mlm_probability 0.3, batch_size 4096, num_steps 30000, rope_base 10000
• MLM-8192: lr 5e-5, mlm_probability 0.3, batch_size 1024, num_steps 30000, rope_base 160000
• WSC: max_len 512, lr 5e-5, batch_size 28672, num_steps 100000
• Fine-tuning: TODO

Evaluation

MTEB

The results of other models are retrieved from MTEB leaderboard.

The gte evaluation setting: mteb==1.2.0, fp16 auto mix precision, max_length=8192, and set ntk scaling factor to 2 (equivalent to rope_base * 2).

Model Name	Param Size (M)	Dimension	Sequence Length	Average (56)	Class. (12)	Clust. (11)	Pair Class. (3)	Reran. (4)	Retr. (15)	STS (10)	Summ. (1)
gte-large-en-v1.5	409	1024	8192	65.39	77.75	47.95	84.63	58.50	57.91	81.43	30.91
mxbai-embed-large-v1	335	1024	512	64.68	75.64	46.71	87.2	60.11	54.39	85	32.71
multilingual-e5-large-instruct	560	1024	514	64.41	77.56	47.1	86.19	58.58	52.47	84.78	30.39
bge-large-en-v1.5	335	1024	512	64.23	75.97	46.08	87.12	60.03	54.29	83.11	31.61
gte-base-en-v1.5	137	768	8192	64.11	77.17	46.82	85.33	57.66	54.09	81.97	31.17
bge-base-en-v1.5	109	768	512	63.55	75.53	45.77	86.55	58.86	53.25	82.4	31.07

LoCo

Model Name	Dimension	Sequence Length	Average (5)	QsmsumRetrieval	SummScreenRetrieval	QasperAbastractRetrieval	QasperTitleRetrieval	GovReportRetrieval
gte-qwen1.5-7b	4096	32768	87.57	49.37	93.10	99.67	97.54	98.21
gte-large-v1.5	1024	8192	86.71	44.55	92.61	99.82	97.81	98.74
gte-base-v1.5	768	8192	87.44	49.91	91.78	99.82	97.13	98.58

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}
}