BugSum-Qwen-7B

BugSum-Qwen-7B is a LoRA adapter for generating concise, developer-facing summaries of software bug reports. It was trained by Md. Abid Shahriar as part of thesis research on bug-report summarization.

This repository contains the adapter and tokenizer files, not a standalone 7B model. It must be loaded on top of Qwen/Qwen2.5-Coder-7B-Instruct.

Intended use

• Research and educational experiments in software-engineering NLP
• Drafting concise summaries from English bug-report descriptions
• Human-assisted issue triage

Generated summaries should be reviewed by a person. The model can omit important details or generate unsupported claims, and it is not intended for automatic safety-critical or production decisions.

Training

The adapter was trained with supervised QLoRA using 4-bit NF4 quantization. The LoRA configuration uses rank 8, alpha 16, dropout 0.05, and targets the attention and MLP projection modules. Training was performed on a consumer NVIDIA RTX 4070 12 GB GPU.

The training corpus is not redistributed in this repository because the source datasets have their own access and licensing conditions.

Loading the adapter

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_id = "Qwen/Qwen2.5-Coder-7B-Instruct"
adapter_id = "Abid-Shahriar/BugSum-Qwen-7B"

tokenizer = AutoTokenizer.from_pretrained(adapter_id)
base_model = AutoModelForCausalLM.from_pretrained(base_id, device_map="auto")
model = PeftModel.from_pretrained(base_model, adapter_id)

Limitations

• Designed for English software bug reports, not general-purpose summarization
• May hallucinate, over-compress, or preserve title-like phrasing instead of producing an ideal summary
• Has not completed a broad external human evaluation
• Results depend on the prompt and decoding settings

Licensing and attribution

The base model is distributed separately under its own Apache-2.0 license. This adapter is currently shared privately for research review while a project-level release license is finalized. Dataset rights are not transferred with this repository.

AI tools assisted parts of the research workflow and documentation; the researcher controlled the data preparation, training, evaluation, and final decisions.