pytorch-training-recipe
Produce an implementable PyTorch or PyTorch Lightning training recipe with dataset handling, optimization, logging, checkpointing, and debugging checks. Use when turning a paper idea into runnable training code — including new experiments, reproductions, fine-tuning setups, scaling recipes, and debugging unstable training.
How do I install this agent skill?
npx skills add https://github.com/jurgendn/agent-skills --skill pytorch-training-recipeIs this agent skill safe to install?
- Gen Agent Trust Hubpass
The skill provides a conceptual framework and template for developing machine learning training scripts using PyTorch. It contains no executable logic, network requests, or credential handling.
- Socketpass
No alerts
- Snykpass
Risk: LOW · No issues
What does this agent skill do?
PyTorch Training Recipe
Translate a research idea into a training setup that can fail loudly, debug quickly, and scale only after correctness is verified.
For concrete implementation checklists and templates, read
references/training-recipe-details.md when needed.
Intake
Specify:
- task, dataset, and split protocol;
- model family and input/target format;
- loss and metrics;
- baseline or reproduction target;
- hardware/compute budget;
- framework preference: raw PyTorch or Lightning.
Workflow
- Define the experiment and claim it supports.
- Define the data pipeline, including leakage checks.
- Choose the simplest training stack that supports the experiment.
- Set optimization choices: optimizer, schedule, batch size, precision, clipping.
- Add first-run safety checks.
- Define metrics and error slices.
- Define logging, checkpoints, seeds, and config recording.
- Flag likely failure modes.
- Recommend a staged rollout from tiny sanity run to full experiment.
Non-Negotiable Checks
- Run a shape/dtype/device check.
- Verify loss is finite at initialization.
- Overfit a tiny batch.
- Run one deterministic seed twice if reproducibility matters.
- Compare against a simple baseline before scaling.
- Save config, commit, seed, and command with each run.
Rules
- Do not start with distributed training.
- Do not tune hyperparameters before the tiny-batch check passes.
- Do not trust a metric until its implementation is checked.
- Keep the first runnable version boring.
Output
Return:
- training recipe;
- data pipeline;
- model/loss/optimizer choices;
- minimal config;
- first-run checks;
- logging/checkpointing plan;
- likely failure modes;
- scale-up plan.
How can the creator link this skill?
Add the canonical catalog link to the repository README so users can inspect current installs and available audits. The publishing guide covers the complete discovery path.
<a href="https://skillzs.dev/skills/jurgendn/agent-skills/pytorch-training-recipe">View pytorch-training-recipe on skillZs</a>