v1.2.2

Improved

• After a crash, the app can now ask for permission to share a crash report. You can preview exactly what will be sent and manage automatic sharing in Settings.
• Imported models now report missing or invalid tokenizer files more clearly.

Fixed

• Fixed some equations in "Under the hood" clipping symbols at their edges.
• Fixed training and inference failing to start on macOS 15 in builds made with newer macOS SDKs.
• Fixed the app icon appearing with the wrong shape on macOS 15.

v1.2.1

Improved

• Direct-download builds now check for updates automatically by default, with controls in Settings to turn automatic checks off or check manually.

Fixed

• Fixed SmolLM2 base-model imports failing because its official tokenizer does not include every standalone byte symbol.
• Fixed additional math rendering bugs in "Under the hood" that lead to crashes. If you're curious, I just couldn't get Swift math libraries to work reliably across macOS versions - so now we're shipping math equations as pre-rendered PDFs.

v1.2.0

New

• Edit, disable, or reset the system prompt in Chat to explore how it changes a model's responses.
• Resolve phone-number labels in imported iMessage conversations from Contacts, entirely on-device.

Improved

• Switch between chapters and pages more quickly.

Fixed

• Fixed chapter math occasionally crashing during page layout or failing to render correctly.
• Fixed crashes caused by diverged training runs and malformed model, tokenizer, checkpoint, or optimizer data.
• Fixed Sampling and Chat carrying stale state between models or checkpoints.
• Fixed several training-resume, validation-cancellation, and checkpoint-consistency edge cases.
• Fixed iMessage imports reading inconsistent database snapshots or leaving temporary data behind after cancellation.
• Fixed storage moves and Reset all data leaving downloaded base models behind.
• Fixed rare Chat truncation around Unicode characters and stop markers.

v1.1.0

New

• Fine-tune open-source base models, including Supra 50M and Qwen 2.5 models.
• Build fine-tuning datasets from iMessage conversations, including group chats and speaker labels.
• Mix multiple datasets and send imported or trained models directly to Sampling.
• Reveal downloaded datasets in Finder.

Improved

• Follow live progress while longer training preparations are running.
• Keep the app more responsive during and after training.
• Scale fine-tuning batches to each model's context and available unified memory, with estimates calibrated from measured MLX peaks.
• Clarify the base-model, importing, and fine-tuning guidance throughout the app.

Fixed

• Fixed a crash in the introductory chapter.
• Fixed lab and chat buttons that could appear enabled but do nothing on macOS 15.
• Fixed several imported-model and training workflow edge cases.

v1.0.0

Notarized direct download for Apple silicon Macs (macOS 15+). Installed apps update automatically via Sparkle.

v1.0.0

v1.0.0+1

Add README