AI Blocks › Build and ship software
Model training recipe finder
Say what you are training. Get the right settings, a safe training loop and a checklist for when the loss goes wrong.
Model training recipe finder is a free AI skill for the job "Build and ship software". Give the instructions to an AI assistant that supports the required files and tools, then review its output.
Free · No account · MIT licence · Made by Orchestra Research · Level: advanced · Thinking: high
Made by Orchestra Research. See the original on GitHub; we copied it as it is, with its licence.
What it does
The right training settings for what you are building, plus a safe loop and a checklist for when the loss misbehaves. It saves the weeks that usually go to trial and error. You start with a table that matches your data type and size to a model: images, text, tables, audio, molecules, proteins or medical scans. The skill then supplies tested patterns for PyTorch: weight setup, optimizer choices, learning rate groups, data loading, checkpoints and spreading work across GPUs. It adds scaling rules for your compute budget and a keep or discard loop for running many experiments. The core idea is that a good recipe beats a fancier model.
You give it
- What you are training and on what data
- Data size
- Your hardware and budget
- Where it fails right now (optional)
- Codebase (optional)
You get
- A model choice with reasons
- Optimizer and learning rate settings
- A tested training loop
- A loss troubleshooting checklist
- Scaling advice for your budget
Use it when
- You are training a model and do not know where to start with settings.
- The loss goes to NaN or stops improving.
- You must pick a model size for a fixed GPU budget.
- You work on biology or medical data and need specific advice.
How to use it
- Add it to your AI once: download it and add the folder to Claude, ChatGPT or Codex as a skill.
- Ask for the job in your own words. Your AI opens the skill when it fits.
- It works best when your AI can run Python code.