---

name: training description: "Route NAVSIM learned-agent training plans, feature and target caches, Hydra configuration, resource selection, and checkpoint diagnosis." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

NAVSIM training

Use this route when a researcher needs to plan or diagnose learned-agent training, feature/target construction, gzip-pickle caches, or the Lightning training wrapper. Keep this route planning-first: do not download data, build large caches, launch training, or run benchmark workloads unless the researcher has explicitly approved the concrete command and resource budget.

Route by need

• Start with workflows.md for the safe sequence from agent selection through cache reuse and checkpoint handoff.
• Read cache-and-builders.md for builder contracts, cache file names, cache-only semantics, and cache compatibility.
• Read configuration.md for Hydra composition, exact override patterns, legal training splits, workers, and CPU/GPU planning.
• Use troubleshooting.md for install/import, backend, data/config, CLI/API, and workflow-specific failures.
• Before any expensive action, run the read-only training-config checker: python scripts/inspect_training_config.py --help, then provide a split and config/overrides as described in its help.

Operating rules

1. Treat the selected train_test_split, log lists, agent configuration, and feature/target builder names as one cache identity. Do not reuse a cache merely because its directory exists.
2. For use_cache_without_dataset=true, require a non-null cache_path and force_cache_computation=false. The runner asserts this combination before constructing CacheOnlyDataset; it is not a performance hint.
3. For training, use a permitted training split such as navtrain, trainval, mini, or navmini. Never train on test, challenge, or private challenge splits; see the legality table in configuration.md.
4. Prefer a small, explicit resource plan. A CPU plan is appropriate for imports and configuration checks, not for claiming TransFuser training parity. The documented GPU defaults use mixed precision and distributed strategy and must be reviewed against available devices.
5. Check the generated output/checkpoint contract before handing a model to an evaluation route. A checkpoint path is an input to learned-agent loading, not a substitute for matching agent configuration and cache builders.