This repository accompanies the preprint of our paper:
ReqBrain: Task-Specific Instruction Tuning of LLMs for AI-Assisted Requirements Generation
arXiv: https://arxiv.org/abs/2505.17632
The paper is under peer review and has not yet been formally published. This repository is made publicly available to promote transparency, reproducibility, and early feedback.
If you use, modify, or build upon any code, data, or models from this repository, please cite the arXiv preprint and this GitHub repository provided below.
@misc{habib2025reqbrain,
title={ReqBrain: Task-Specific Instruction Tuning of LLMs for AI-Assisted Requirements Generation},
author={Mohammad Kasra Habib and Daniel Graziotin and Stefan Wagner},
year={2025},
eprint={2505.17632},
archivePrefix={arXiv},
primaryClass={cs.SE},
url={https://arxiv.org/abs/2505.17632},
}Copyright © Mohammad Kasra Habib, 2025. All rights reserved.
No official open-source license is granted until the paper is formally published. This repository is not considered open-source in its current state. Use is limited to non-commercial, private research purposes only. Commercial use and redistribution are strictly prohibited without prior permission from the first author.
After publication, the repository may be archived on Zenodo with an open license to ensure reproducibility and prevent versioning conflicts.
If you use any part of this repository, please cite the arXiv preprint and this GitHub repository.
Requirements Brain, an open-source, task-specific instruction-tuned language model, is trained to generate authentic and adequate software requirements. It generates requirements to support elicitation and specification. Requirements engineers can interact with ReqBrain in a chat format to generate new requirements, generate requirements and classify on the fly, and turn bug descriptions into well-structured requirements. Additionally, they can provide an existing set of requirements to identify missing ones, classify requirements by type, or combine it with RAG to process huge bulks of proprietary textual data and elicit well-structured requirements in context.
The repository provides:
- Datasets for training and evaluation
- Training Scripts for training ReqBrain
- Evaluation data and scripts to assess performance
- A Jupyter Notebook demo showcasing model usage
- ReqBrain Candidates: Link to trained models and their weights
You can explore the basic functionality of ReqBrain through the interactive Jupyter Notebook demo:
- Model Download: It automatically downloads the model weights from our Hugging Face repository.
- Step-by-Step Guidance: The notebook guides you through the process and explains the key steps involved.
- GPU Requirement: A GPU with a minimum of 32GB GPU memory is required to ensure smooth execution.
Download the reqbrain_demo.ipynb script from the repo’s root directory, then launch it using:
jupyter notebook reqbrain_demo.ipynb
The training_scripts directory provides scripts for fine-tuning ReqBrain, organized in Jupyter Notebooks:
- Model-Specific Scripts: Each notebook is named after the model it fine-tunes, making it easy to locate.
- Step-by-Step Guidance: Users are guided through fine-tuning using comments.
- GPU Requirement: A GPU with a minimum of 32GB GPU memory is required to ensure smooth execution.
The five trained models are quite large. For easy access, seamless downloading, further tuning, and integration with other Hugging Face tools, they are hosted on ReqBrain's Hugging Face page. The link to each trained model is provided below:
| Training Script | Name on HuggingFace | HuggingFace Model Link |
|---|---|---|
training_scripts/train_falcon-base.ipynb |
ReqBrain-falcon-7b | https://huggingface.co/kasrahabib/ReqBrain-falcon-7b |
training_scripts/train_falcon-instruct.ipynb |
ReqBrain-falcon-7b-instruct | https://huggingface.co/kasrahabib/ReqBrain-falcon-7b-instruct |
training_scripts/train_llama2.ipynb |
ReqBrain-Llama-2-7b-chat-hf | https://huggingface.co/kasrahabib/ReqBrain-Llama-2-7b-chat-hf |
training_scripts/train_mistralai.ipynb |
ReqBrain-Mistral-7B-Instruct-v0.2 | https://huggingface.co/kasrahabib/ReqBrain-Mistral-7B-Instruct-v0.2 |
training_scripts/train_zephyr.ipynb |
※ReqBrain-zephyr-7b-beta | https://huggingface.co/kasrahabib/ReqBrain-zephyr-7b-beta |
※ Evaluation results indicate that this model achieves the highest performance across all tasks.
All evaluation data and scripts can be found in the evaluations directory. ReqBrain is evaluated through:
- Automated NLP Metrics: Located in evaluations/automated_nlp_evaluations/
- Human Evaluators: Located in evaluations/human_evaluations/
Further details are provided inside each of the subdirectories: /evaluations/automated_nlp_evaluations/ and evaluations/human_evaluations/.
The dataset directory contains the instruct dataset.
- Training Set is located in
dataset/train/ - Evaluation Set: is located under
dataset/test/
The dataset is structured to be compatible and directly used with Falcon, LLaMA, Zephyr, and other LLMs sharing a similar input format.
Make sure you have the datasets library installed. You can do this using pip:
pip install datasetsWe provide a helper function, get_dataset_by_model_format(), to easily filter the dataset based on the target model format.
import datasets
def get_dataset_by_model_format(dataset, split, ds_format):
return dataset[split].filter(lambda example: example['ds_format'] == ds_format)# Load the dataset
dataset_path = './path_to_the_dataset/dataset'
instruct_dataset = datasets.load_from_disk(dataset_path)
# Filter the dataset by the desired format (e.g., 'falcon')
dataset = get_dataset_by_model_format(instruct_dataset, split='train', ds_format='falcon')
dataset_test = get_dataset_by_model_format(instruct_dataset, split='test', ds_format='falcon')Replace falcon with llama or zephyr as needed to retrieve the dataset for the appropriate model.