A working local fitting path
The MVP fits a separate Gemma 3 1B IT four-bit MLX LoRA adapter on Apple Silicon. A fixed recipe uses one LoRA layer, rank 8, eight steps, batch size 1, and a 128-token sequence limit. Original checkpoints remain unchanged. Supported hardware and model-format compatibility are checked before work begins.
Permission is an execution boundary
Datasets require explicit training consent and rights confirmation. Training grants bind the model, dataset revision and digest, recipe, wall-time and memory budgets, and experiment count. The fixed supervisor checks these records and launches a child with OS network denial and constrained file access. The trusted supervisor itself is not fully filesystem-sandboxed. Cancellation, timeout, interrupted-job recovery, and export cleanup have been exercised.
A functional result, not a quality leaderboard
An original fixture with 16 training, 4 validation, and 4 test examples completed actual fitting and held-out evaluation. Mean answer-token cross entropy was 13.0833 for the base and approximately 1.9292 with the adapter. Four test examples establish a functional path, not broad generalization, production quality, or a comparison against an API model.
Why lower loss cannot approve a model
A deliberately conflicting-label experiment also reduced held-out loss: 9.2964 to 4.6712. It did not produce the intended regression rejection. This unsuccessful hypothesis is retained as evidence that loss alone is insufficient. Product evaluation records remain inactive and explicitly disallow activation.
The remaining lifecycle
Behavioral regression tests, compatible adapter inference, explicit owner approval and activation, and rollback remain under development. MLX adapter compatibility with GGUF inference is not assumed. The Fine-tuning Agent is currently a constrained experiment service, not an autonomous self-improving model.
Evidence & reading copy
Download the scoped study or engineering note. Figures are descriptive; they do not run the product in your browser.
local-adaptation.md Full note · Markdown