---

name: training description: "Guide Boltz training and retraining through config edits, debug launches, resource settings, checkpoints, and training troubleshooting." disable-model-invocation: true metadata: disco-role: operating license: MIT

Boltz Training

Use this sub-skill when the user wants to train, retrain, resume, debug, or modify Boltz training configs for structure or confidence models.

Route First

• Raw-data preprocessing, RCSB/OpenFold processing, MSA generation, clustering, CCD, Redis, or mmseqs setup belongs in the data-preparation sub-skill.
• Prediction outputs, confidence interpretation, benchmark aggregation, or post-training evaluation belongs in prediction or evaluation.
• Boltz docs state that updated Boltz-2 training information is coming soon; do not invent unreleased Boltz-2 training recipes.

Start Here

1. Confirm the user already has processed training data, processed MSAs, and a ligand symmetry file; full public preprocessed data is about 250 GB before any training outputs.
2. Choose a config shape from references/training-configs.md: structure, confidence-only, or full structure-plus-confidence.
3. Validate placeholders, paths, split files, checkpoint intent, and resource settings with scripts/boltz_training_config_check.py.
4. Run a debug launch before any full run; debug mode disables multi-device DDP, sets data workers to zero, and disables wandb in the training script.
5. Treat full training as expensive and hardware-dependent; never promise that it is cheap, quick, or feasible without suitable GPUs and storage.

Common Workflows

• Training launch sequence, debug overrides, DDP notes, wandb behavior, and checkpoint decisions: references/training-workflows.md.
• Config fields, template differences, data module requirements, split semantics, and resource knobs: references/training-configs.md.
• Failure diagnosis for placeholders, missing data, Hydra imports, GPU memory, wandb, DDP, and Boltz-2 docs gaps: references/troubleshooting.md.

Native Check

Run the bundled static checker from this sub-skill directory before a launch:

python scripts/boltz_training_config_check.py path/to/train.yaml --repo-root . --profile debug

Use --check-imports only inside a Boltz training environment where the package and training dependencies are installed.