InstructUIE

InstructUIE

• This repo releases our implementation for the InstructUIE model.
• It is built based on the pretrained Flan T5 model, and finetuned on our data (IE INSTRUCTIONS).
• If you are looking for NER datasets or models, you may also refer to our recent work: B2NER. It provides a standardized and comprehensive bilingual (Chinese-English) NER data collection, suitable for LLM training.

Requirements

Our main experiments and analysis are conducted on the following environment:

• CUDA (11.3)
• cuDNN (8.2.0.53)
• Pytorch (1.10.0)
• Transformers (4.26.1)
• DeepSpeed (0.7.7)

You can install the required libraries by running

bash setup.sh

Data

Our models are trained and evaluated on IE INSTRUCTIONS. You can download the data from Baidu NetDisk or Google Drive.

Training

A sample script for training the InstructUIE model in our paper can be found at scripts/train_flan-t5.sh. You can run it as follows:

bash ./scripts/train_flan-t5.sh

Released Checkpoints

We have released our 11B UIE model, click here for download.

Evaluation

A sample script for evaluating the InstructUIE model in our paper can be found at scripts/eval_flan-t5.sh. You can run it as follows:

bash ./scripts/eval_flan-t5.sh

The decoded results would save to predict_eval_predictions.jsonl in your output dir. To calculate f1 with predict_eval_predictions.jsonl

python calculate_f1.py

Citation

@article{wang2023instructuie,
  title={InstructUIE: Multi-task Instruction Tuning for Unified Information Extraction},
  author={Wang, Xiao and Zhou, Weikang and Zu, Can and Xia, Han and Chen, Tianze and Zhang, Yuansen and Zheng, Rui and Ye, Junjie and Zhang, Qi and Gui, Tao and others},
  journal={arXiv preprint arXiv:2304.08085},
  year={2023}
}