cli-v0.554.0

Highlights

• SFT context parallelization: Training now processes per-sample rows as fragments in SFT context parallel mode, improving efficiency when handling large samples. Sample packing remains disabled with documented reasoning for the RL gaps.
• Changelog corrections: Restored TRL phase-2 entries that were inadvertently moved, ensuring accurate release history for changes that landed after v0.552.0.

Full changelog: https://github.com/prbe-ai/probe/blob/main/CHANGELOG.md