AI Blocks › Set up AI to work for you
AI model training debugger
Describe your model, data and what goes wrong. Get a check order that finds silent bugs before tuning.
AI model training debugger is a free AI skill for the job "Set up AI to work for you". Give the instructions to an AI assistant that supports the required files and tools, then review its output.
Free · No account · Made by AI Blocks · Level: expert · Thinking: high
What it does
Model training fails quietly. Code runs, numbers come out, and the model is still wrong because of a bad label, a mixed-up shape or a wrong setting. People then tune settings for days on top of a bug. This skill sets an order of work. It starts with looking through the data by hand for a long time, then a small model that is hard to break, with a fixed seed and a dumb baseline. It checks the starting loss against what chance predicts, tries to overfit one tiny batch, grows the model one piece at a time, and only then adds limits to reduce overfitting. Before every step it states the expected result, so a surprise tells you something.
You give it
- The model, the task and the data, with sizes
- What you see, such as loss curves or scores (optional)
- Your training code or settings (optional)
You get
- The order of checks for your case, with the expected result for each
- What each failed check points to
- A list of changes made one at a time with results
- Where to stop checking and start tuning
Use it when
- Your model trains without errors and still gives poor results.
- You are about to tune settings and are not sure the pipeline is right.
- A model that worked before has stopped improving.
How to use it
- Start a chat with the skill. Open it if you added it to your AI, or copy it and paste it in.
- Describe the model, the task, the data and what goes wrong. Paste the loss curve or scores if you have them.
- Do the first check it gives and send back what you saw.
- Make only one change per run and write down the result.
- Move to tuning only when all the checks pass.