SyncBench / README.md
xuehang's picture
Update README.md
d8f9577 verified
metadata
license: mit

SyncBench

In tackling the challenge of agent out-of-sync, SyncBench is released as a part of our work SyncMind: Measuring Agent Out-of-Sync Recovery In Collaborative Software Engineering

Dataset Description

SyncBench for agent out-of-sync recovery:

  • Task: Agent out-of-sync
  • SyncBench
    • In our current version, we construct SyncBench based on 21 popular GitHub repositories, while our dataset construction approach allows adaptive upsampling and downsampling for custom needs
      • Expand to incorporate other repositories as data sources
      • Downsampled to smaller ealuation subsets
  • Datasets: Caller and Callee
    • Caller: Testing code out-of-sync
    • Callee: Tested code out-of-sync

Data Structure

  • Format: JSON and CSV

  • Fields:

    • instance_id: SyncBench instance ID
    • instance_type: data type (i.e., caller or callee)
    • task_type: task type (i.e., fail-to-pass or pass-to-pass)
    • repo_url: URL link of the source GitHub repository
    • commit: Git commit ID
    • fm_type: Function type (i.e., function or method)
    • fm_name: Function name
    • original_code: Out-of-sync function code
    • gold_code: Up-to-date function code
    • old_filtered_context: Out-of-sync context code (filtered)
    • old_complete_context: Out-of-sync context code (complete)
    • new_filtered_context: Up-to-date context code (filtered)
    • new_complete_context: Up-to-date context code (complete)
    • initial_error_log: Initial execution error log
    • original_summary: Parsed execution output of the out-of-sync code
    • gold_summary: Parsed execution output of the up-to-date code
    • pyfile_name: Name of the Python file
    • pyfile_path: Path to the Python file
    • unittest_path: Path to the unit test
  • Dataset Versions

    • SyncBench
      • Data: syncbench_24k.csv
      • Size: 24,332
    • Evaluation Dataset
      • Data: syncbench_300.csv
        • Callee: syncbench_callee_150.csv
        • Caller: syncbench_caller_150.csv
      • Size: 300

Usage

from datasets import load_dataset
dataset = load_dataset("xuehang/SyncBench")