agentleFS
jsboigejsboige/CoursIA/.claude/skills/train-model/SKILL.md

train-model

Skill17 starsChanged yesterdayMITView on GitHub

Train an ML model in the QuantConnect ML-Training-Pipeline with thermal-safe GPU usage and rigorous validation. Arguments: <architecture|script> [--dry-run] [--seeds 0,1,7,42,99] [--folds 5] [--bg]

The file

We split this skill into the rules it gives an agent

1 rule. About 0 are common, 0 are distinctive, and 1 is very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈0

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 1

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Check a intervalles utiles : tail -50 <log> | grep -E "FINAL|RESULT|ERROR|temp".

    line 42

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Train Model

Entrainer un modele ML du pipeline MyIA.AI.Notebooks/QuantConnect/ML-Training-Pipeline/ avec securite thermique GPU et validation rigoureuse. Couvre l'Epic #1454 (Training & Post-Training : RL/PPO, Decision Transformer, GenAI fine-tuning/LoRA).

Pour les cycles longs/batch, deleguer a l'agent training-specialist en async (run_in_background: true).

Arguments

  • <architecture|script> : nom court (lstm, transformer, mamba, patchtst, itransformer, moe, gnn, dqn-rl, decision-transformer) ou nom de script train_*.py.
  • --dry-run : smoke test CPU via validate_training_package.py --verbose (le script n'accepte pas --dry-run ; ses flags réels sont --verbose et --script — validate_training_package.py:148-149), sans GPU. Toujours commencer par la.
  • --seeds 0,1,7,42,99 : seeds pour le multi-seed (defaut >=4 parmi 0/1/7/42/99).
  • --folds 5 : walk-forward folds (defaut 5).
  • --bg : lancer le training en background (recommande pour les longs runs).

Process

Phase 0 — Environnement + grounding

  1. Activer l'env : & "C:\Users\MYIA\miniconda3\envs\coursia-ml-training\python.exe".
  2. Lire REGISTRY.md + CURRICULUM.md : le modele est-il deja un keeper valide ? deprecated ? Ne pas refaire.

Phase 1 — Dry-run CPU (HARD avant tout GPU)

python validate_training_package.py --verbose  # smoke test, pas de GPU (le script n'accepte pas --dry-run)

Si le dry-run echoue : reparer (pas contourner). Ne pas passer en GPU sur un package casse.

Phase 2 — Verifier le watchdog thermique (HARD)

python -c "from gpu_training import batch_thermal_check, get_gpu_temp; print('temp:', get_gpu_temp())"
  • Temperature non nulle => watchdog actif (MAX_TEMP=80C, cool_sleep=30, AMP).
  • ImportError ou temp 0 => protection thermique en no-op silencieux (gpu_training.py absent). NE PAS lancer un long run GPU sans protection : restaurer le module ou monitorer nvidia-smi manuellement. Documenter.

Phase 3 — Training

  • Lancer python scripts/train_<arch>.py (via env coursia-ml-training), en --bg pour les longs runs.
  • Noter l'ID du BG + nature attendue. Pendant le run, avancer une 2e track (2 tracks min, cf CLAUDE.md operations longues).
  • Check a intervalles utiles : tail -50 <log> | grep -E "FINAL|RESULT|ERROR|temp". Pas event-par-event.

Phase 4 — Validation rigoureuse (gate PR)

  • Walk-forward 5-fold (WalkForwardSplitter), pas single split.
  • >=4 seeds ; edge >= 2sigma cross-seed, sinon flag "noise".
  • DM significance (dm_test.py / diebold_mariano.py).
  • Comparaison majority_class_baseline + couts de transaction (5bps SPY, 10bps crypto).
  • Pas de FAANG/Mag7 en training set.
  • Verdict honnete : BEATS / NO BEATS / INCONCLUSIVE. Jamais "promising". Single-seed => [POC] explicite.

Phase 5 — Registry + livrable

  • python registry_update.py pour mettre a jour REGISTRY.md.
  • Checkpoints .pt volumineux/transients : ne pas stager dans la PR.
  • Rapport (metriques par seed/fold, verdict, DM p-value) sur le dashboard, pas dans un fichier du repo.
  • PR : un sujet, body avec WF + seeds + edge + baseline + tx costs + verdict (sinon CHANGES_REQUESTED).

Anti-patterns interdits

  • Sauter le dry-run et lancer direct en GPU.
  • Lancer un long run GPU sans verifier le watchdog importable (Phase 2).
  • Claim "BEATS" sans multi-seed + DM significance.
  • Python systeme au lieu de coursia-ml-training.
  • Stager des .pt lourds.

Voir aussi

  • .claude/agents/training-specialist.md — orchestrateur async
  • docs/archive/ml-trading-state.md — anti-FAANG, multi-seed, walk-forward
  • .claude/rules/pr-review-discipline.md C — gate ML PR

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.