gte-modernbert-base

We are excited to introduce the gte-modernbert series of models, which are built upon the latest modernBERT pre-trained encoder-only foundation models. The gte-modernbert series models include both text embedding models and rerank models.

The gte-modernbert models demonstrates competitive performance in several text embedding and text retrieval evaluation tasks when compared to similar-scale models from the current open-source community. This includes assessments such as MTEB, LoCO, and COIR evaluation.

Model Overview

• Developed by: Tongyi Lab, Alibaba Group
• Model Type: Text Embedding
• Primary Language: English
• Model Size: 149M
• Max Input Length: 8192 tokens
• Output Dimension: 768

Model list

Usage

[!TIP] For transformers and sentence-transformers, if your GPU supports it, the efficient Flash Attention 2 will be used automatically if you have flash_attn installed. It is not mandatory.

pip install flash_attn

Use with transformers

# Requires transformers>=4.48.0

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-modernbert-base"
tokenizer = AutoTokenizer.from_pretrained(model_path)
model = AutoModel.from_pretrained(model_path)

# 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())
# [[42.89073944091797, 71.30911254882812, 33.664554595947266]]

Use with sentence-transformers:

# Requires transformers>=4.48.0
from sentence_transformers import SentenceTransformer
from sentence_transformers.util import cos_sim

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

model = SentenceTransformer("Alibaba-NLP/gte-modernbert-base")
embeddings = model.encode(input_texts)
print(embeddings.shape)
# (4, 768)

similarities = cos_sim(embeddings[0], embeddings[1:])
print(similarities)
# tensor([[0.4289, 0.7131, 0.3366]])

Use with transformers.js:

// npm i @huggingface/transformers
import { pipeline, matmul } from "@huggingface/transformers";

// Create a feature extraction pipeline
const extractor = await pipeline(
  "feature-extraction",
  "Alibaba-NLP/gte-modernbert-base",
  { dtype: "fp32" }, // Supported options: "fp32", "fp16", "q8", "q4", "q4f16"
);

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

// Compute similarity scores
const similarities = (await matmul(embeddings.slice([0, 1]), embeddings.slice([1, null]).transpose(1, 0))).mul(100);
console.log(similarities.tolist()); // [[42.89077377319336, 71.30916595458984, 33.66455841064453]]

Additionally, you can also deploy Alibaba-NLP/gte-modernbert-base with Text Embeddings Inference (TEI) as follows:

• CPU

docker run --platform linux/amd64 \
  -p 8080:80 \
  -v $PWD/data:/data \
  --pull always \
  ghcr.io/huggingface/text-embeddings-inference:cpu-1.7 \
  --model-id Alibaba-NLP/gte-modernbert-base

• GPU

docker run --gpus all \
  -p 8080:80 \
  -v $PWD/data:/data \
  --pull always \
  ghcr.io/huggingface/text-embeddings-inference:1.7 \
  --model-id Alibaba-NLP/gte-modernbert-base

Then you can send requests to the deployed API via the OpenAI-compatible v1/embeddings route (more information about the OpenAI Embeddings API):

curl https://0.0.0.0:8080/v1/embeddings \
  -H "Content-Type: application/json" \
  -d '{
    "input": [
      "what is the capital of China?",
      "how to implement quick sort in python?",
      "Beijing",
      "sorting algorithms"
    ],
    "model": "Alibaba-NLP/gte-modernbert-base",
    "encoding_format": "float"
  }'

Training Details

The gte-modernbert series of models follows the training scheme of the previous GTE models, with the only difference being that the pre-training language model base has been replaced from GTE-MLM to ModernBert. For more training details, please refer to our paper: mGTE: Generalized Long-Context Text Representation and Reranking Models for Multilingual Text Retrieval

Evaluation

MTEB

The results of other models are retrieved from MTEB leaderboard. Given that all models in the gte-modernbert series have a size of less than 1B parameters, we focused exclusively on the results of models under 1B from the MTEB leaderboard.

LoCo (Long Document Retrieval)(NDCG@10)

COIR (Code Retrieval Task)(NDCG@10)

BEIR(NDCG@10)

Hiring

We have open positions for Research Interns and Full-Time Researchers to join our team at Tongyi Lab. We are seeking passionate individuals with expertise in representation learning, LLM-driven information retrieval, Retrieval-Augmented Generation (RAG), and agent-based systems. Our team is located in the vibrant cities of Beijing and Hangzhou. If you are driven by curiosity and eager to make a meaningful impact through your work, we would love to hear from you. Please submit your resume along with a brief introduction to dingkun.ldk@alibaba-inc.com.

Citation

If you find our paper or models helpful, feel free to give us a cite.

@inproceedings{zhang2024mgte,
  title={mGTE: Generalized Long-Context Text Representation and Reranking Models for Multilingual Text Retrieval},
  author={Zhang, Xin and Zhang, Yanzhao and Long, Dingkun and Xie, Wen and Dai, Ziqi and Tang, Jialong and Lin, Huan and Yang, Baosong and Xie, Pengjun and Huang, Fei and others},
  booktitle={Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: Industry Track},
  pages={1393--1412},
  year={2024}
}

@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}
}