Training loop debugger
Skill maximussthegreat/ml-researcher-os/skills/training-loop-debugger
Agent skills and workflows for reproducible ML research.
npx -y skills add maximussthegreat/ml-researcher-os --skill training-loop-debuggerAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its author says it does
Copied from the file, not written here
Use when reviewing ML training code, suspicious metrics, unstable loss, broken evaluation, or a model that appears too good to be true.
SKILL.md
1.5 KB, as published. Nobody here has run it
Training Loop Debugger
Use this skill to review training and evaluation code before trusting results.
Goal
Find the smallest code or protocol issue that could invalidate reported model metrics.
Review order
- Data loading
- Split creation
- Preprocessing fit and transform
- Model initialization
- Loss and optimizer
- Training mode and eval mode
- Metric computation
- Checkpoint selection
- Test set use
- Logging and reproducibility
High-risk bugs
Immediately flag:
- fitting scalers, tokenizers, imputers, or feature selectors on validation or test data
- using test data for early stopping
- computing metrics on logits when probabilities or labels are required
- forgetting
model.eval()during evaluation - forgetting
torch.no_grad()during evaluation - using different preprocessing at train and eval time
- reporting the best validation run as test performance
- changing seeds until a good result appears
Required output
| Area | Finding | Severity | Evidence | Fix |
|---|
Severity must be one of:
blockerhighmediumlow
Rules
- Do not claim the model is fixed until a rerun is available.
- Do not rewrite the whole project when a small patch explains the issue.
- If logs are missing, ask for logs or mark the conclusion as provisional.