Release Highlights
What's Changed
- Add tool support for RULER evaluation (#542)
- Remove SkyPilot backend and docs (#541)
- refactor: streamline AutoTrajectoryContext management (#540)
- feat: group-level trajectory metadata (#539)
- Correct Qwen3 30B model name and link (#538)
- Document Qwen3 30B (#537)
- Vendor tinker_cookbook under art package (#535)
- feat: replace pyright with ty for type checking (#534)
- feat(tinker): add TinkerNativeBackend (#532)
- fix: align Tinker server port with client base_url (#531)
- Fix model registration (#529)
- fix: preserve tool-call context in tokenization (#527)
- Fix OpenAI temperature handling (#526)
- SFT preprocessing (#525)
- Remove dev/playwright_agent directory (security + cleanup) (#524)
- feat: add persistent state methods to Model (#522)
- feat: Backend-First Training API (Phase 1) (#521)
- Remove TorchTune service and related dev projects (#520)
- feat: move trajectory logging from backend to frontend (#518)
- feat: Replace pyright with ty for type checking (#517)
- feat: Replace run_checks.sh with prek pre-commit hooks (#516)
- Multi-checkpoint inference for pipelined training (RFC #513) (#515)
- Use training_step for W&B x-axis to allow out-of-order logging (#512)
- Update wandb and weave (#507)
- Upgrade litellm, suppress bugs (#506)
- bump: Release v0.5.7 (#505)
Full Changelog: v0.5.7...v0.5.9