AGENT SKILL

training-data

Plan DiffDock score/confidence training and prepare datasets, splits, ESM embeddings, caches, and checkpoints safely.

License 许可
MIT
最近更新
2026-10-08
Stars
★ 11
Forks
3

01 前言 Front Matter

name
training-data
license
MIT
metadata
{ "disco-role": "operating" }
description
Plan DiffDock score/confidence training and prepare datasets, splits, ESM embeddings, caches, and checkpoints safely.
disable-model-invocation
true

02 资源文件 4

03 SKILL.md 正文

DiffDock Training Data

Use this sub-skill when a user wants to plan DiffDock score-model training, confidence-model training, dataset layout checks, split/caching decisions, ESM embedding preparation, or checkpoint compatibility before running expensive jobs.

Start Here

Safe Helpers

Scope And Routing

This sub-skill covers training planning and data preparation only: score-model and confidence-model training flags, dataset families, split files, ESM embedding preparation/conversion, cache naming, model-parameter outputs, and checkpoint compatibility.

Route these tasks elsewhere:

  • Use trained checkpoints for docking prediction, inference YAML tuning, CSV/batch inference, and confidence-guided sampling: ../docking-inference/SKILL.md.
  • Aggregate benchmark metrics, run post-training evaluation, or compare PoseBusters/DockGen/PDBBind results: ../evaluation-benchmarks/SKILL.md.
  • Launch or debug the Gradio UI: ../web-ui/SKILL.md.

Quick Commands

python sub-skills/training-data/scripts/build_training_command.py --mode score --dataset pdbbind --data-dir data/PDBBind_processed --split-train data/splits/timesplit_no_lig_overlap_train --split-val data/splits/timesplit_no_lig_overlap_val --run-name pdbbind_trial --limit-complexes 16 --n-epochs 1 --batch-size 2

python sub-skills/training-data/scripts/validate_dataset_layout.py --dataset-type pdbbind --dataset-root data/PDBBind_processed --split-path data/splits/timesplit_no_lig_overlap_train --max-complexes 5

python sub-skills/training-data/scripts/validate_esm_embedding_index.py data/esm2_embeddings.pt --expect-ids data/splits/timesplit_no_lig_overlap_train

Key Safety Rules

  • Treat training, confidence-dataset generation, ESM extraction, graph preprocessing, and evaluation as expensive GPU/storage jobs; do not start them without explicit user approval.
  • Use the bundled helpers first: they print commands or JSON checks and never run DiffDock training.
  • Keep dataset paths, cache paths, and run directories relative to the active DiffDock checkout or user-provided project workspace.
  • Do not assume full imports are available: full training paths require the optional-heavy Torch, PyG, e3nn, W&B, RDKit, ProDy, ESM/OpenFold stack and preferably CUDA.
  • Preserve model_parameters.yml with checkpoints; downstream confidence training and inference compatibility depend on it.

04 同仓库更多 Skills