Gemma 4 TFD Agentic Trajectories Dataset for Agent Software Engineering Trajectories

Dataset Overview

Gemma 4 TFD Agentic Trajectories is an execution trajectory and benchmark evaluation dataset for agentic software engineering tasks. Built around the Google Gemma 4 model, it aims to evaluate the performance of different agent strategies in code debugging and repair tasks by recording the execution process and test feedback when agents solve software engineering problems.

This dataset is built from tasks adapted from SWE-bench Lite. It contains 50 software engineering problems and provides multi-turn agent execution trajectories, tool-call records, automated test validation results, and benchmark evaluation data for different agent strategies.

Dataset Composition

• Agent execution trajectories: records the multi-turn interactions, task execution, and code repair process of agents solving software engineering problems, presenting the execution steps from problem localization to repair validation;
• Tool-call records: include tool-call information during agent execution, covering steps such as code search, file operations, and test execution, which can be used to analyze the agent's operational workflow and execution strategy;
• Test feedback and validation results: provide automated unit-test feedback and validation information to assess whether code modifications meet task requirements and the agent's subsequent correction process based on test results;
• Benchmark evaluation data: through files such as benchmark_summary.csv, provide task evaluation results for different agent strategies, including standard ReAct and test-feedback-driven (TFD-Agent) strategies, for comparing task resolution rates and execution efficiency;
• Evaluation and reproduction resources: provide agent trajectory files, benchmark evaluation data, and related documentation to support research on agentic software engineering tasks, strategy comparison, and experiment reproduction.