Skip to content

v1.0.0-alpha.29

Pre-release
Pre-release

Choose a tag to compare

View all tags
@stainless-app stainless-app released this 26 Feb 22:05
v1.0.0-alpha.29
110b294
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

1.0.0-alpha.29 (2026-02-26)

Full Changelog: v1.0.0-alpha.28...v1.0.0-alpha.29

Features

  • add client-side max seq length filter and align defaults (cbac95f)
  • training shape auto-config, DAPO/GSPO losses, deployment readiness probing (4d1f333)

Bug Fixes

  • align DAPO/GSPO loss behavior with reference implementations (4f65e1e)
  • align TIS with slime -- add icepop, safe ratio, default clip=2.0 (2acdcc3)
  • clarify weight sync variable naming in cookbook recipes (f213b88)
  • orthogonal TIS, deployment parallelization, stale CISPO refs (0a90ecb)
  • pass return_dict=False to apply_chat_template for transformers>=5 compat (5a660aa)
  • resolve pyright lint errors in train_sft.py (586f1c1)
  • return_dict compatibility, training shape error messages, GSPO PPO clipping (928a09e)
  • skip shape-derived fields in trainer API when using training shapes (72e0bcc)
  • sort imports to pass ruff I001 lint (0d3d8ec)
  • trim redundant and over-mocked tests (df3060e)

Chores

  • internal: make test_proxy_environment_variables more resilient to env (6de8f74)

Refactors

  • simplify TIS to vanilla implementation with safety clamp (7f2d7f1)
Assets 2
Loading