Experiments and text training

MLX Workbench is the native SwiftUI interface for the shared engine. Agent integrations can use mlx-agent lab with the same database, artifacts, jobs, and review records.

Runtime

The core agent remains dependency-light. Experiments require Python 3.12 and requirements-lab.txt. Optional local MPS video generation uses requirements-lab-inference.txt. Install requirements into the selected interpreter; model weights must be present locally.

On macOS, the default experiment database and artifacts live under ~/Library/Application Support/mlx-agent/lab/. Explicit --database and --artifacts select another workspace. Standalone scripts/video-lab preserves legacy JSON output.

python scripts/mlx-agent lab --help
python scripts/mlx-agent lab experiment list --json

Heavy generation, conversion, and training share a private compute lease. A live owner blocks conflicting launches. Receipt-backed detached jobs retain ownership; dead processes can be reclaimed, while unverifiable live identity fails closed.

Native workflows

Compare → Controlled studies creates a one-variable frozen comparison, queues generation, and retains artifacts. Blind review uses a common playback timeline. Complete rubric scores and a preference are required before immutable submission and authorized reveal. Adaptive studies preserve their budgets and progression.

Workspace… previews adoption into a new destination. Stop workers and the former web application first. Adoption verifies database and artifact content, preserves IDs and sealed assignments, and retains source backups. It refuses active work, changed inputs, or an occupied destination.

Activity reconciles legacy conversion receipts before starting an imported pending item. Import requires a reviewed content digest and retains the source and a backup.

Text LoRA

Training previews validate every bounded JSONL record and identify dataset content and the cached base model. Confirmation rechecks the preview and copies the dataset into a private run snapshot. Each job retains its own receipt, log, actual exit result, and verified adapter content identity.

A completed adapter can be evaluated against its pinned base model using identical prompt, seed, token limit, and sampling settings. Separate processes produce baseline and adapter outputs. The engine rechecks inputs before accepting a content-addressed report into the experiment store. Native Training exposes these retained evaluations.

An existing directory alone does not establish successful training. Historical receipts without exit evidence remain unknown; receipts without pinned base identity cannot authorize a new linked evaluation. Comparison outputs are evidence and do not automatically record a quality judgment.

Media training remains governed by its existing qualification gates. Text LoRA acceptance does not qualify video training.