CoderForge-Preview: SOTA open dataset for training efficient coding agents

Together AI · 4 months ago

A research team released CoderForge-Preview, an open dataset containing 258,000 test-verified coding agent trajectories spanning 51,000 tasks across 1,655 repositories. A Qwen 32B model fine-tuned on this data achieved 59.4% pass@1 on SWE-Bench Verified, ranking first among open-data models in the ≤32B parameter range. The dataset enables open-source researchers to train and improve coding agents without relying on proprietary data sources.