training-orchestrator

name	training-orchestrator
description	Orchestrate personal model training — SFT, DPO, and STT fine-tuning via mlx-tune on Apple Silicon.
tags	["training","self-improvement"]
metadata	{"author":"fae","version":"2.0"}

Training Orchestrator

This skill manages the personal model training pipeline using mlx-tune — a native MLX fine-tuning library with Unsloth-compatible API.

Training Modes

Mode	Script	Purpose
SFT	train	Supervised fine-tuning on conversation episodes
DPO	train_dpo	Preference learning from user corrections
STT	train_stt	Speech-to-text adaptation for Qwen3-ASR
Keyword	train_keyword	Barge-in keyword classifier (micro 1D-CNN)
Speech Verifier	train_speech_verifier	Speech/music/noise classifier (micro 1D-CNN)

Available Scripts

• export_data: Export conversation episodes to train.jsonl/valid.jsonl (80/20 split)
• train: SFT fine-tuning via mlx-tune SFTTrainer (Qwen3.5 models)
• train_dpo: DPO preference training via mlx-tune DPOTrainer (uses correction pairs)
• train_stt: STT fine-tuning via mlx-tune FastSTTModel (Qwen3-ASR)
• evaluate: Benchmark trained checkpoint — reads train_metrics.json or falls back to log parsing
• propose: Generate human-readable comparison report for the user
• check_status: Check if training is in progress
• deploy: Activate a trained adapter (requires user approval)
• rollback: Revert to previous adapter
• train_keyword: Train 5-class keyword classifier (interrupt/wake/speech/silence/noise)
• train_speech_verifier: Train 3-class speech verifier (speech/music/noise)

Model Map

Preset	Model	RAM Requirement
tiny	Qwen3.5-2B-4bit	<16 GB
small	Qwen3.5-4B-4bit	≥16 GB
medium	Qwen3.5-9B-unsloth-mlx	≥32 GB
large	Qwen3.5-35B-A3B-4bit	≥48 GB

Auto-selection matches Fae's production LLM model tier.

SFT Workflow

1. Run export_data to prepare training dataset from recent conversations
2. Run train with preset (smoke/light/standard/deep)
3. Run evaluate to benchmark the new checkpoint
4. Run propose to generate upgrade proposal
5. If user approves, run deploy to switch models

DPO Workflow

1. Run training-data-bridge extract_corrections to build DPO pairs
2. Run train_dpo with preset (smoke/light/standard)
3. Run evaluate to benchmark
4. Run propose → deploy if approved

STT Workflow

1. Fae automatically captures ASR corrections ("my name is X not Y")
2. Run train_stt to fine-tune Qwen3-ASR with correction pairs
3. Run evaluate → propose → deploy

Training Presets (SFT)

Preset	Steps	Batch	Grad Accum	LR	LoRA Rank	Seq Length
smoke	10	1	1	1e-4	8	512
light	50	2	2	5e-5	16	1024
standard	200	4	4	2e-5	16	2048
deep	500	4	4	1e-5	32	2048

Safety

• Training data never leaves the Mac — 100% local via MLX
• New models must score ≥ current on evaluation
• User must explicitly approve model switches
• Previous checkpoints always preserved for rollback
• mlx-tune runs in uv sandbox with PEP 723 inline dependencies