v0.2.47 1.10.0
data-privacy-stack/presidioSep 3, 2026 → Sep 4, 20261 day between releases
What's New
Significant performance improvements and behavior-altering defaults in version 1.10.0.
Breaking Changes
- Validation now evaluates one model per epoch instead of two
- grad_accum_steps defaults to 1 instead of 4
- Optimizer parameter groups collapsed, affecting positional indexing
- Dataset builders now strictly validate config instead of falling back to defaults
New Features
- GPU batched linear-assignment solver for CUDA
- Performance optimizations for training and inference
Improvements
- Up to 46% faster inference on large frames
- 44% faster training time
- Optimized per training step speed via optimizer parameter-group merge
- DataLoader throughput increased by 1.25–2.26x
Upgrade Notes
Update training scripts to accommodate the new default grad_accum_steps of 1 and the collapsed optimizer parameter groups. Ensure dataset configurations are explicitly defined as validation now strictly enforces config settings.
The Two Releases
Open Recorder v0.2.47
Sep 3, 2026· 2 months ago
Minimal release notes: "## Commits since v0.2.46 - de8bafb Merge pull request #1080 from imbhargav5/codex/native-shortcut-recorder - 7729870 fe"
Full release notesv1.10.0: Faster Training & Inference
Sep 4, 2026· 2 months ago
Focuses on significant performance improvements for training and inference, accompanied by several behavior-altering defaults.
Full release notes