---

name: train-evaluate description: "Run and debug Lightning-Hydra-Template training and evaluation commands, checkpoints, callbacks, loggers, accelerators, and safe repeated runs." disable-model-invocation: true metadata: disco-role: operating license: MIT

Train and Evaluate

Use this sub-skill when the user wants to run, resume, debug, or adapt the template's Lightning training/evaluation workflow.

Triggers

Read this sub-skill for tasks about:

• python src/train.py, python src/eval.py, train_command, or eval_command.
• Checkpoint resume/evaluation using ckpt_path, last.ckpt, or ModelCheckpoint output.
• Callbacks, loggers, logger=csv, W&B/Neptune/Comet/MLflow/Aim configs, or hyperparameter logging.
• Trainer presets such as trainer=cpu, trainer=gpu, trainer=ddp, trainer=ddp_sim, and trainer=mps.
• Fast-dev, batch-limit, overfit, or profiler training commands.
• Repeated training commands or adapting the repo's scheduling shell pattern.

Quick workflow

1. Inspect CLI/config without training:

   train_command --help
   eval_command --help
   python <this-skill>/scripts/check_lightning_hydra_project.py --repo-root . --config-name train.yaml --instantiate
   

2. For a safe smoke run, disable online loggers and use a debug config. Remember the MNIST example can download data:

   python src/train.py debug=fdr logger=null
   

3. For normal training, choose experiment, trainer, logger, and tags explicitly:

   python src/train.py experiment=example trainer=cpu logger=csv tags='[mnist,baseline]'
   

4. For resume, pass the checkpoint to the train config:

   python src/train.py ckpt_path=/path/to/checkpoints/last.ckpt
   

5. For evaluation, pass a checkpoint to the eval config:

   python src/eval.py ckpt_path=/path/to/checkpoints/last.ckpt logger=null
   

6. To preview repeated runs without executing them, use scripts/run_scheduled_training.sh.

Read references

• Training and evaluation: train/eval object flow, checkpoint semantics, commands, metrics, and safe validation.
• Loggers, callbacks, and accelerators: logger packages/credentials, callback defaults, trainer groups, and backend scope.
• Troubleshooting: missing checkpoint, monitor metric, logger credentials, MNIST/data download, DDP, batch divisibility, and Hydra output paths.

Bundled script

• scripts/run_scheduled_training.sh: safe adaptation of the repository's scripts/schedule.sh; defaults to dry-run and requires --execute before launching training.

Boundaries

• For Hydra defaults, experiment YAML, sweeps, and debug config composition, use configure-experiments.
• For adding new data/model code or fixing _target_ imports, use customize-data-model.
• For pytest/CI choices or no-network smoke gates, use test-maintain-template.