Skip to content

ISE-Research/girt-model

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

6 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

GIRT

HuggingFace Model HuggingFace Demo HuggingFace Dataset

TL;DR

The repository introduces GIRT-Model, an open-source assistant language model that automatically generates Issue Report Templates (IRTs) or Issue Templates. It creates IRTs based on the developer’s instructions regarding the structure and necessary fields.

Links

How to load model (local)

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

# load model and tokenizer
model = AutoModelForSeq2SeqLM.from_pretrained('nafisehNik/girt-t5-base')
tokenizer = AutoTokenizer.from_pretrained('nafisehNik/girt-t5-base')

# Ensure that the model is on the GPU for cpu use 'cpu' instead of 'cuda'
model = model.to('cuda')


# method for computing issue report template generation
def compute(sample, top_p, top_k, do_sample, max_length, min_length):

    inputs = tokenizer(sample, return_tensors="pt").to('cuda')

    outputs = model.generate(
        **inputs,
        min_length= min_length,
        max_length=max_length,
        do_sample=do_sample,
        top_p=top_p,
        top_k=top_k)

    generated_texts = tokenizer.batch_decode(outputs, skip_special_tokens=False)
    generated_text = generated_texts[0]
    
    replace_dict = {
        '\n ': '\n',
        '</s>': '',
        '<pad> ': '',
        '<pad>': '',
        '<unk>!--': '<!--',
        '<unk>': '',
    }
    
    postprocess_text = generated_text
    for key, value in replace_dict.items():
        postprocess_text = postprocess_text.replace(key, value)

    return postprocess_text

prompt = "YOUR INPUT INSTRUCTION"
result = compute(prompt, top_p = 0.92, top_k=0, do_sample=True, max_length=300, min_length=30)

Dataset

A dataset in the format of pairs of instructions and corresponding outputs. GIRT-Instruct is constructed based on GIRT-Data, a dataset of IRTs. We use both GIRT-Data metadata and the Zephyr-7B-Beta language model to generate the instructions. This dataset is used to train the GIRT-Model.

We have 4 different types in GIRT-Instruct. These types include:

  • default: This type includes instructions with the GIRT-Data metadata.
  • default+mask: This type includes instructions with the GIRT-Data metadata, wherein two fields of information in each instruction are randomly masked.
  • default+summary: This type includes instructions with the GIRT-Data metadata and the field of summary.
  • default+summary+mask: This type includes instructions with the GIRT-Data metadata and the field of summary. Also, two fields of information in each instruction are randomly masked.

How to load dataset

from datasets import load_dataset
dataset = load_dataset('nafisehNik/girt-instruct', split='train')
print(dataset['train'][0]) # First row of train

Code

The code for fine-tuning the GIRT-Model and evaluation in a zero-shot setting is available here. It downloads the GIRT-Instruct and fine-tunes the t5-base model.

We also provide the code and prompts used for the Zephyr model to generate summaries of instructions.

UI (online)

This UI is designed to interact with GIRT-Model, it is also accessible in huggingface: https://huggingface.co/spaces/nafisehNik/girt-space

  1. IRT input examples
  2. metadata fields of IRT inputs
  3. summary field of IRT inputs
  4. model config
  5. generated instruction based on the IRT inputs
  6. generated IRT

GIRT

Citation

This work is accepted for publication in MSR 2024 conference, under the title of "GIRT-Model: Automated Generation of Issue Report Templates".

@inproceedings{nikeghbal2024girt-model,
  title={GIRT-Model: Automated Generation of Issue Report Templates},
  booktitle={21st IEEE/ACM International Conference on Mining Software Repositories (MSR)},
  author={Nikeghbal, Nafiseh and Kargaran, Amir Hossein and Heydarnoori, Abbas},
  month={April},
  year={2024},
  publisher={IEEE/ACM},
  address={Lisbon, Portugal},
  url = {https://doi.org/10.1145/3643991.3644906},
}