Awesome

An Out-of-the-box Large Language Model for Open Domain Sequence Understanding

Spotlights

• A bilingual model (English and Chinese) specially enhanced for open-domain NLU.
• Trained with diverse synthesized data and high-quality NLU dataset.
• Handle all NLU tasks that can be transformed into a combination of atomic tasks, classification and extraction.

📰 Update News

SeqGPT is continuously updating. We have provided online demos for everyone. In the future, we will provide new versions of models with upgraded capabilities. Please continue to pay attention!

• [2023/10/09] 💪 We provide API of SeqGPT-3B for users who want to access larger SeqGPT.
• [2023/09/20] 🎛️ We provide a sample script for fine-tuning on a custom dataset at here.
• [2023/08/23] 🛠️ We release the weight of SeqGPT-560M at both Modelscope and Hugging Face. You can download and inference with our model simply following the usage case.
• [2023/08/23] 🔥 We provide an online demo of SeqGPT at Modelscope! Try it now!
• [2023/08/21] 📑 We release the paper of SeqGPT: SeqGPT: An Out-of-the-box Large Language Model for Open Domain Sequence Understanding. More implementation details and experimental results are presented in the paper.

Performance

We perform a human evaluation on SeqGPT-7B1 and ChatGPT using the held-out datasets. Ten annotators are tasked to decide which model gives the better answer or two models are tied with each other. SeqGPT-7B1 outperforms ChatGPT on 7/10 NLU tasks but lags behind in sentiment analysis (SA), slot filling (SF) and natural language inference (NLI).

Usage

Install

conda create -n seqgpt python==3.8.16

conda activate seqgpt
pip install -r requirements.txt

Inference

from transformers import AutoTokenizer, AutoModelForCausalLM, AutoModel
import torch

model_name_or_path = 'DAMO-NLP/SeqGPT-560M'
tokenizer = AutoTokenizer.from_pretrained(model_name_or_path)
model = AutoModelForCausalLM.from_pretrained(model_name_or_path)
tokenizer.padding_side = 'left'
tokenizer.truncation_side = 'left'

if torch.cuda.is_available():
    model = model.half().cuda()
model.eval()
GEN_TOK = '[GEN]'

while True:
    sent = input('输入/Input: ').strip()
    task = input('分类/classify press 1, 抽取/extract press 2: ').strip()
    labels = input('标签集/Label-Set (e.g, labelA,LabelB,LabelC): ').strip().replace(',', '，')
    task = '分类' if task == '1' else '抽取'

    # Changing the instruction can harm the performance
    p = '输入: {}\n{}: {}\n输出: {}'.format(sent, task, labels, GEN_TOK)
    input_ids = tokenizer(p, return_tensors="pt", padding=True, truncation=True, max_length=1024)
    input_ids = input_ids.to(model.device)
    outputs = model.generate(**input_ids, num_beams=4, do_sample=False, max_new_tokens=256)
    input_ids = input_ids.get('input_ids', input_ids)
    outputs = outputs[0][len(input_ids[0]):]
    response = tokenizer.decode(outputs, skip_special_tokens=True)
    print('BOT: ========== \n{}'.format(response))

Citation

If you found this work useful, consider giving this repository a star and citing our paper as followed:

@misc{yu2023seqgpt,
      title={SeqGPT: An Out-of-the-box Large Language Model for Open Domain Sequence Understanding}, 
      author={Tianyu Yu and Chengyue Jiang and Chao Lou and Shen Huang and Xiaobin Wang and Wei Liu and Jiong Cai and Yangning Li and Yinghui Li and Kewei Tu and Hai-Tao Zheng and Ningyu Zhang and Pengjun Xie and Fei Huang and Yong Jiang},
      year={2023},
      eprint={2308.10529},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}