Automatic Training and LoRA
Canonical source:
docs/AUTOMATIC_TRAINING.md. This page is refreshed from the Finch repository when the website builds.
Status: Deferred. Automatic training and LoRA adapter loading are not supported Finch features.
Finch does not automatically collect OpenAI-compatible requests, start a training timer or worker, invoke Python, rewrite a training queue, generate a LoRA adapter, or hot-load adapters.
Explicit feedback submitted with Ctrl+G/Ctrl+B, feedback commands, or
POST /v1/feedback is stored in ~/.finch/feedback.jsonl. The store is private
(~/.finch mode 0700 and the file mode 0600 on Unix), append-only, locked across
processes, and synced before acknowledgement. Feedback records do not consent
to or trigger training.
Existing ~/.finch/training_queue.jsonl files and adapter files are preserved
without processing, deletion, or migration. The legacy Python setup command is
an explicit manual experiment and is not connected to the daemon or feedback
path. Finch disabled that background Python path in
#139 (disable automatic Python LoRA training unless a supported native path is explicitly enabled).
Why this work is deferred
Section titled “Why this work is deferred”The retained code is scaffolding, not a usable training feature. Completing the path currently requires all of the following unsupported pieces:
- a way to apply adapters in Finch’s ONNX inference runtime, tracked in #1 (LoRA adapter loading at ONNX runtime);
- an external training stack such as MLX on Apple Silicon or PyTorch and PEFT on Linux/CUDA, plus adapter conversion and a memory-bounded reload strategy, tracked in #7 (LoRA training memory efficiency); and
- a verified model/runtime compatibility matrix, tracked in #74 (refresh the local-model catalog and publish a runtime compatibility matrix).
Finch will use pre-trained artifacts and provider-backed models instead. Training should not be presented as upcoming or partially available unless the project adopts a supported path with explicit privacy, resource, cancellation, recovery, and retention controls.