README_EN.md

Long Bert Chinese

English | 简体中文

Long Bert: A long text similarity model that supports a length of 8192 tokens. Based on bert-base-chinese, it changes the original BERT positional encoding to ALiBi positional encoding, allowing BERT to support a sequence length of 8192.

News

• Support CoSENT fine-tuning. see train folder
• The model has been uploaded to Huggingface

Usage

from numpy.linalg import norm
from transformers import AutoModel
model_path = "OctopusMind/longbert-8k-zh"
model = AutoModel.from_pretrained(model_path, trust_remote_code=True)

sentences = ['我是问蚂蚁借呗为什么不能提前结清欠款', "为什么借呗不能选择提前还款"]
embeddings = model.encode(sentences)
cos_sim = lambda a,b: (a @ b.T) / (norm(a)*norm(b))
print(cos_sim(embeddings[0], embeddings[1]))

CoSENT fine-tuning data format

[
    {
        "sentence1": "一个男人在吹一支大笛子。",
        "sentence2": "一个人在吹长笛。",
        "label": 3
    },
    {
        "sentence1": "三个人在下棋。",
        "sentence2": "两个人在下棋。",
        "label": 2
    },
    {
        "sentence1": "一个女人在写作。",
        "sentence2": "一个女人在游泳。",
        "label": 0
    }
]

Fine-tuning

cd train/

python cosent_finetune.py \
        --data_dir ../data/train_data.json \
        --output_dir ./outputs/my-model \
        --max_seq_length 1024 \
        --num_epochs 10 \
        --batch_size 64 \
        --learning_rate 2e-5

Contributions

Feel free to make contributions to this module by submitting pull requests or raising issues in the repository.

License

This project is licensed under the Apache-2.0 License.