mbtorch-examples

What problem does it solve?

Examples and demos can easily drift out of sync with evolving APIs, become too large to run, or require heavy external setup that prevents users from quickly verifying behavior. This skill provides a lightweight, repeatable approach to designing, implementing, and maintaining examples so that they remain minimal, executable, and informative.

Core Features & Use Cases

• Minimal runnable examples: Guidelines to create one-purpose examples with required files (main.mbt or lib.mbt, moon.pkg, README) that run with the moon toolchain and avoid internal/private APIs.
• Example types and conventions: Patterns for CLI training scripts, browser/WASM demos, and edge-simulation examples; naming rules and a difficulty ladder from Hello-MLP to LoRA-style fine-tuning.
• Maintainability & testing: Advice to keep examples executable with synthetic data, include quickstart steps, and add lightweight tests so CI can verify buildability and basic runtime behavior.
• Use case: When adding a new fine-tuning API, create examples/fine_tune_lora demonstrating a minimal training loop, a README with the run/build steps, and a small test that executes a few training steps.

Quick Start

Create a minimal example directory that includes a main.mbt or lib.mbt, a moon.pkg, and a README showing the single moon run or moon build command needed to execute the demo.