v0.2.0
Released April 2, 2026.
- Added
SWEBenchProEnvironmentfor repo-level patch tasks backed by ScaleAI/SWE-bench_Pro - Added
TerminalBenchEnvironmentfor shell-task evaluation backed by Terminal-Bench 2.0 - Added
MixedEnvironmentfor ratio-based multi-benchmark reward routing through existing TRL/reward APIs - Added reusable
PatchVerifierfor unified-diff parsing, apply checks, and repo test scoring - Extended reward integrations so per-sample benchmark metadata can flow through existing batch interfaces