training toolkit · compatible toolkit

MLX-LM LoRA / QLoRA

An explicitly started local fine-tuning command using GoldCap-prepared data on compatible Apple silicon systems.

Troubleshoot this path

Curious Obmil studies an observatory instrument beside a smaller happy GoldCap with one main cap, two small caps, and living gold veins — Curious checks the signal. GoldCap keeps the local path close to the user.

DecaCap interface: Local command launched by the companion
Published evidence: GoldCap can prepare canonical data and the companion can launch a configured MLX command. Successful training depends on the local installation, model, memory, data, and command profile.
Exact-model probe: Not published Account-, model-, region-, quota-, runtime-, and hardware-specific verification still requires a real user-owned key or device.
Last reviewed: 2026-07-26

REQUIREMENTS

Apple silicon Mac
Compatible Python environment
mlx-lm installed locally
A supported base model
Enough unified memory and storage

LIMITATIONS

DecaCap does not guarantee model/toolkit compatibility
A finished adapter is not automatically activated
Training quality is not inferred from process completion

IMPLEMENTATION TEST TARGETS

Explicit start
Observable process status
Pause/resume/stop where supported
Adapter path reporting
Runtime retest boundary

WHAT THIS DOES NOT PROVE

Access to a particular model on a user account
Quality, speed, pricing, quota, or safety behavior
Hardware fit or training-toolkit compatibility