daVinci-Dev: Agent-native Trajectories for Software Engineering from GitHub Pull Requests

Agent-native trajectories used in the daVinci-Dev project for mid-training in software engineering. The dataset includes trajectories constructed from GitHub pull requests, specifically a Python variant. The dataset was created by GAIR and was last updated on the platform in January 2026.

Use Cases

• Training AI agents for code review and patch generation based on pull request trajectories.
• Studying developer workflows and agent behavior in software engineering contexts.
• Benchmarking agent performance on tasks derived from real-world GitHub repositories.
• Developing models for automated software maintenance and bug fixing.

Strengths

• Data is derived from real-world GitHub pull requests, providing practical context.
• Focuses on agent-native trajectories, a specific and relevant data type for AI agent training.
• Includes a Python-specific variant of the data.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

GAIR

Collection Method

Constructed from GitHub pull requests.

Time Range

null

Freshness

Last updated 2026-01-31 05:45:31; freshness should be verified.

Geography

null