---

name: train-configure description: "Configure DreamerV3 training, evaluation, logging, checkpoints, and smoke runs from the CLI." disable-model-invocation: true metadata: disco-role: operating license: MIT

Train Configure

Use this sub-skill when you need to:

• compose DreamerV3 config blocks from dreamerv3/configs.yaml
• launch, resume, or dry-run dreamerv3.main.main
• choose between train, train_eval, eval_only, parallel, and the parallel_* worker modes
• inspect startup flow, logging, checkpointing, and resume behavior
• run a bounded CPU or random-agent smoke before heavier GPU training

Read these bundled references first:

• references/cli-and-config.md
• references/run-loops.md
• references/troubleshooting.md

Run the bundled smoke helper when you want a one-command CPU validation:

• python scripts/smoke_train_debug.py --help
• python scripts/smoke_train_debug.py --dry-run-config
• python scripts/smoke_train_debug.py

Covered surface

This sub-skill covers the control and configuration layer around the DreamerV3 training entry point:

• dreamerv3/main.py
  • main
  • make_agent
  • make_env
  • make_replay
  • make_stream
  • make_logger
• dreamerv3/configs.yaml
  • defaults
  • size presets
  • task presets
  • debug
  • multicpu
  • CLI override ordering
• embodied/run/train.py
• embodied/run/train_eval.py
• embodied/run/eval_only.py
• embodied/run/parallel.py
• README install/training/tips distilled into runtime guidance

What to do

1. Start from the config reference and pick the smallest preset set that matches the goal.
2. Use the smoke helper for a bounded CPU or random-agent check before any larger run.
3. Keep --logdir stable when you want continuation, but start a fresh logdir after any model-shape change.
4. Use the run-loop reference to choose the correct script mode and checkpoint path.
5. If the issue is env wrappers, replay internals, or stream mechanics, hand off to embodied-dataflow.
6. If the issue is RSSM/encoder/decoder/model internals, hand off to jax-models.
7. If the issue is Docker, plotting, or backend ops, hand off to results-ops.

Fast rules

• --configs blocks apply in order, then explicit CLI flags win.
• debug is for fast validation, not for learning a good policy.
• random_agent is the cheapest way to check env, replay, logger, and checkpoint plumbing.
• train_ratio is interpreted relative to batch size and batch length, so changing those changes the effective update cadence.
• task must resolve to a supported suite/task pair or env creation will fail before training starts.
• Reusing a checkpoint after changing a size preset or other model-shape-sensitive config is unsafe unless compatibility is known.

Expected outputs

A healthy bounded training smoke should usually produce at least:

• config.yaml
• metrics.jsonl
• scores.jsonl
• ckpt/
• terminal metrics and episode summaries

For parallel modes, expect a nested checkpoint tree such as ckpt/agent, ckpt/replay, and ckpt/logger.

When to stop here

Do not dig into source-repo implementation details beyond the bundled references if you only need to:

• choose a config preset
• verify a small training or evaluation smoke
• resume from a checkpoint in a compatible logdir
• decide whether a failure belongs to config, logging, checkpointing, or run-loop control