Skip to content

v0.2.2

Choose a tag to compare

View all tags
@YujiaBao YujiaBao released this 02 Apr 19:33
· 105 commits to main since this release
v0.2.2
41bccbf
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

v0.2.2

New distillation recipes, multi-domain RL support, and Kimi K2 weight merging.

New features

  • SDFT (Self-Distillation Fine-Tuning) recipe with top-K distillation (#524)
  • Off-policy top-K distillation for multi-teacher knowledge merging (#572)
  • Kimi K2 / K2.5 shard-by-shard merge with INT4 expert dequant/requant (#573)
  • InterleavedRLDatasetBuilder for multi-domain RL training (#570)
  • 22 marimo tutorials (101–503) (#562)

Bug fixes

  • Fix tokenizer_class corrupted to TokenizersBackend during export (#582)
  • Fix experts-fp8 compression_config for non-DeepSeek models (#580)
  • Fix InterleavedRLDataset crash on ragged last source batch (#574)
  • Other minor bug fixes (#567, #577)

Other

  • Extend SFT LR sweep to full Tinker model lineup (#575)
  • Make image_processor optional for Qwen3VL renderers (#566)

See the full CHANGELOG.md for details.

Assets 2
Loading