What it is
Builds infrastructure that converts scientific code repositories into executable, verifiable agent environments, then uses the resulting trajectories to train scientific coding models. Instead of treating scientific repos as static benchmarks, turns them into reusable executable environments with expert-calibrated scientific checks, task factories, and verifier rewards for SFT, RL, and eval.
Key results
- 64 environments from 27 codebases with 2,812 validated tasks
- ScienceIDE-Hard has 85 tasks across 18 environments for evaluation
- Best reported ScienceIDE-Hard success rate is 67.1% with Claude Fable 5.1
Scientific code can be packaged into verifiable environments that train agents and transfer gains beyond the original scientific tasks.