daVinci-Dev: Agent-native Mid-training for Software Engineering
- Type
- dataset
- Venue
- GAIR (Global AI Research Lab) / SII
- Year
- 2026
- Source
- huggingface
- Access
- restricted
- Language
- English, Python
- Added
- 2026-07-17T20:18:03.726361+00:00
- Verified
- 2026-07-17T20:18:03.726361+00:00
Summary
daVinci-Dev is a dataset of agent-native trajectories for software engineering mid-training, consisting of two complementary sources: ~4.1M LLM-enhanced GitHub pull request trajectories (contextually-native) and test-passing executable rollouts from SWE-Agent + GLM-4.6 on SWE-rebench (environmentally-native). It was used to train the daVinci-Dev-72B and daVinci-Dev-32B models for agentic coding tasks.
Keywords
code software-engineering agent pull-request trajectory synthetic swe-agent mid-training
Topics
Code / Software Engineering
Research notes
- Requires accepting conditions and providing legal name/organization to access. PR content enhanced with Qwen3-235B-A22B-Instruct. Paper: arxiv 2601.18418.