GitHub - FudanSELab/LinuxFLBench (original) (raw)

This repository contains the code and data for the paper "LinuxFLBench: Benchmarking and Enhancing LLM-based Agents in Localizing Linux Kernel Bugs".

Dataset Introduction

LINUXFLBENCH is a new benchmark of 250 Fault Localization tasks derived from real-world Linux kernel bugs.

• The dataset is located at dataset/LINUXFLBENCH_dataset.jsonl in JSON Lines format.
• Each line is a real Linux kernel bug sample, with fields including:
  • id: Bug ID
  • title: Bug title
  • description: Detailed bug description
  • Kernel Version: The version of the Linux kernel in which the bug occurred (e.g., 5.6.7).
  • patch: Patch content for the fix
  • paths: Source file paths involved (i.e., localization target files)
  • methods: Function names involved
  • Additional metadata: kernel version, component, hardware, etc.
• The dataset covers various kernel versions and is suitable for evaluating LLM/agent-based fault localization in large and complex systems(i.e., the Linux kernel).
• The source code for different Linux kernel versions can be downloaded from here.

Methods and Code Structure

The main code is under the code/ directory, organized as follows:

• scale/: Candidate file expansion and reasoning
  • scaling_candidates_with_dir.py: Directory-based candidate expansion
  • scaling_candidates_with_guess.py: LLM-based candidate expansion
• merge/: Multi-method result fusion and reranking
  • merge.py: Fusion of multiple ranking results
  • rerank.py: LLM-based candidate reranking
• mail/：Mail-related scripts
  • mails_retrieval.py：Retrieves relevant emails from the mail dataset based on queries
  • search_mails_bm25s.py：BM25-based Mail Search Utils
• method_fl/: Method-level fault localization based on the predicted code files
  • method_localize.py: Method-level fault localization script
• eval/: Evaluation and metrics
  • evaluate.py: Main evaluation script
  • evaluation_metrics.py: Common metrics such as Recall@K, MRR
• utils.py, file_parser.py: General utility functions
• The mail data for retrieval can be downloaded from here.

Typical Workflow

1. Candidate Expansion
   Use scripts in scale/ to expand candidate file lists for each bug (e.g., Directory-Aware Expansion, Potential Cause Expansion).
2. Candidate Integration
   Use scripts in merge/ to fuse multiple candidate ranking results, and rerank with LLM.
3. Evaluation
   Use scripts in eval/ to evaluate the final results with metrics such as Recall@K and MRR.

Results

All experimental results are located in the result/ directory and can be used for reproduction.

Requirements

This project requires Python 3.8+ and the following packages:

• openai
• jsonlines

Install dependencies with pip:

pip install openai jsonlines

Some scripts require configuration of OpenAI API Key and base_url. See script arguments for details.

Quick Start

Example: Directory-Aware Expansion

python code/scale/scaling_candidates_with_dir.py
--data_path dataset/LINUXFLBENCH_dataset.jsonl
--save_path results/dir_scaling.jsonl
--gpt_base_url https://api.openai.com/v1
--api_key YOUR_API_KEY
--kernel_path /path/to/linux/kernel/

Evaluate the results:

python code/eval/evaluate.py --path results/dir_scaling.jsonl

For more details, usage, or questions, please open an issue or contact the authors.