Run any Skill in Manus with one click

Get Started

training-orchestrator

Stars3

Forks0

UpdatedJuly 2, 2026 at 12:35

Orchestrate personal model training — SFT, DPO, and STT fine-tuning via mlx-tune on Apple Silicon.

Installation

Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.

Run Skill in Manus

Source

saorsa-labs

saorsa-labs/fae

View GitHub Repository View Creator Repositories

Download

Run Skill in Manus

Related occupationsSOC

Based on SOC occupation classification

Data ScientistsComputer and Mathematical Occupations·SOC 15-2051

File Explorer

15 files

SKILL.md

readonly

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

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

DPO Workflow

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

STT Workflow

Fae automatically captures ASR corrections ("my name is X not Y")
Run train_stt to fine-tune Qwen3-ASR with correction pairs
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

More from this repository

same repository

collaborate

saorsa-labs/fae

Collaborate over x0x — contacts, messages, spaces, boards, files, presence, trusted-machine port forwards, and replicated stores. Use for sharing cards, connecting people, or cross-device work.

2026-07-153

capability-discovery

saorsa-labs/fae

Guide for surfacing one relevant Fae capability that the user hasn't set up yet. Warm, specific, one thing at a time.

2026-07-153

self-diagnostic

saorsa-labs/fae

Deep pipeline diagnostic — every model, system specs, audio config, speaker state. Flags missing or broken components. Use for diagnose, debug, or health check.

2026-07-153

training-data-bridge

saorsa-labs/fae

Extract training signals from Fae's memory — SFT examples, DPO correction pairs, engagement scores, and interest-weighted sampling.

2026-07-153

cloud-brain-setup

saorsa-labs/fae

Set up (or turn off) a cloud brain for harder questions via OpenRouter — privacy-first, plain language. Use when the user wants a bigger or smarter brain or mentions the cloud.

2026-07-053

first-launch-onboarding

saorsa-labs/fae

A warm first conversation for someone meeting Fae for the first time — getting to know their name, where they live, and what they care about, showing one thing live, and explaining how to talk to her.

2026-07-053

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

Run export_data to prepare training dataset from recent conversations

Run train with preset (smoke/light/standard/deep)

Run evaluate to benchmark the new checkpoint

Run propose to generate upgrade proposal

If user approves, run deploy to switch models

DPO Workflow

Run training-data-bridge extract_corrections to build DPO pairs

Run train_dpo with preset (smoke/light/standard)

Run evaluate to benchmark

Run propose → deploy if approved

STT Workflow

Fae automatically captures ASR corrections ("my name is X not Y")

Run train_stt to fine-tune Qwen3-ASR with correction pairs

Run evaluate → propose → deploy

Training Presets (SFT)

Preset

Steps

Batch

Grad Accum

LoRA Rank

Seq Length

smoke

1e-4

512

light

5e-5

1024

standard

200

2e-5

2048

deep

500

1e-5

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