Software Agent Execution Trajectories for Tool-Calling Tasks

3,872 complete agent execution trajectories are provided for real-world software engineering tasks. The data is converted into a reasoning and tool-call format with validated state transitions. It was created by AmanPriyanshu and last updated in March 2026.

Use Cases

• Fine-tuning language models to generate tool-call sequences from reasoning traces for code tasks.
• Analyzing agent state transition patterns within the validated FSM structure of the trajectories.
• Training models on multi-language code tasks using trajectories derived from 942 repositories across 10 programming languages.
• Benchmarking agent performance on 3,005 distinct software engineering task instances.

Strengths

• Contains 3,872 complete agent execution trajectories.
• Derived from a base of 3,005 task instances across 942 repositories.
• Trajectories are cleaned and rectified into a strict reasoning and tool-call format.

Limitations

• Specific column names and data schema are not provided.
• The original sample size of 3,005 task instances may be limited for some large-scale training applications.
• Potential class or task distribution imbalances are not detailed.

Provenance

Source

Derived from ernie-research/MEnvData-SWE-Trajectory.

Collection Method

Full agent execution records converted into a reasoning and tool-call format.

Freshness

Last updated 2026-03-07.