M-Courtyard v0.4.9 — Local LoRA fine-tuning on Apple Silicon (MLX)
New in this release:
Custom JSONL Import (D-11): Import your own pre-formatted JSONL datasets directly from the UI. Supports {"prompt", "completion"} and {"messages": [...]} (chat) formats. Backend validates the first 5 lines for field completeness before accepting. If valid.jsonl is absent, the app auto-generates it from train.jsonl at training start.
rsLoRA Scaling Strategy (T-3): New advanced option under the LoRA/DoRA training method. Switches from standard alpha/r to alpha/√r scaling — the approach introduced in Kalajdzievski 2023. Injects use_rslora: true into the lora_config.yaml passed to mlx_lm lora. Best for rank ≥ 32.
Built with: Tauri 2 + React + mlx-lm | macOS 14+ / Apple Silicon only
