Skip to content

v0.0.4

Choose a tag to compare

View all tags
@adohe adohe released this 03 Jul 07:18
v0.0.4
31edcd2

AReno v0.0.4

v0.0.4 focuses on training usability, SFT dataset support, algorithm test coverage, performance cleanup, and Docs V1 polish.

Features and API

  • Support dataset loaders for SFT training (#115, #117)
  • Add an SFT validation error for RL-shaped prompt+solutions datasets (#74)
  • Accept prompt+completion as an SFT dataset alias (#73)
  • Add a GSM8K SFT dataset loader (#71)
  • Add explicit training parameter tuning for rollout-based trainers (#113, #114)
  • Add nightly build and integration tests for core algorithms (#123, #124)

Performance and Fixes

  • Reduce PPO score_logprobs memory usage (#125, #126)
  • Clean up native attention diagnostics and fallback details (#84, #118)
  • Harden packed_next_token_logprobs against zero-length segments (#83)

Documentation

  • Reorganize Docs V1 navigation and structure (#106, #127)
  • Add a Dataset Formats reference guide (#70)
  • Document the current observability surface (#45, #122)

Cleanup and Tests

  • Protect API imports from engine-heavy imports (#65, #121)
  • Add registry discovery smoke coverage (#53, #120)
  • Assert math dataset_loader output is RL-shaped, not SFT-shaped (#72)
Assets 2
Loading