메뉴
Use when a model won't learn, loss is NaN, metrics look too good/bad, or training is unstable. Provides a systematic decision tree for diagnosing data, optimization, and generalization failures.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
| name | ml-debugging |
| description | Use when a model won't learn, loss is NaN, metrics look too good/bad, or training is unstable. Provides a systematic decision tree for diagnosing data, optimization, and generalization failures. |
ML bugs are usually silent — the code runs, but the model is wrong. Debug systematically by isolating which layer is broken: data, optimization, or generalization. Form a hypothesis, test it with the cheapest possible experiment, then move on.
The fastest sanity check in ML. Take a single small batch and train until loss ≈ 0.
x, y = next(iter(train_loader))
for _ in range(200):
optimizer.zero_grad()
loss = criterion(model(x), y)
loss.backward(); optimizer.step()
print(loss.item()) # should approach 0
| Symptom | Likely cause | Fix |
|---|---|---|
| Loss = NaN | LR too high; log(0); /0; bad input scaling | Lower LR, clip grads, add eps, normalize inputs |
| Loss flat at start | LR too low; dead ReLUs; wrong loss | Raise LR; check init; verify loss/target shapes |
| Train great, val terrible | Overfitting or leakage | Regularize/augment; audit for leakage |
| Val better than train | Leakage; val too easy; dropout-at-train artifact | Re-check split; inspect val set |
| Metric implausibly high | Target leakage | Hunt leaked features (see EDA skill) |
| Stuck at chance | Labels misaligned; data not shuffled; LR off | Verify label mapping; shuffle; sweep LR |
refund_issued) used to predict the outcome?[1e-5 ... 1e-1] (log scale) — wrong LR causes most failures.A root-caused fix plus a regression check (the one-batch test + a held-out metric) so experiment-tracking can confirm the issue stays fixed.
Use when preparing raw data for modeling — handling missing values, duplicates, inconsistent types, outliers, and bad categorical values. Emphasizes fitting all imputation on train-only to avoid leakage.
Use when running ML experiments that need to be compared, reproduced, or shared. Covers MLflow/Weights & Biases logging, what to track, run organization, and model registry basics.
Use when starting on a new dataset, before modeling, or when asked to "explore", "profile", "understand", or "summarize" data. Covers structured EDA, distributions, correlations, target leakage checks, and visualization.
Use when creating, encoding, scaling, or selecting features for ML models. Covers categorical encoding, numeric transforms, datetime/text/aggregation features, and leakage-safe target encoding.
Use when optimizing model hyperparameters. Covers search strategy (random vs Bayesian/Optuna), leakage-safe tuning inside CV, search-space design, early stopping, and budget management.
Use when the target is rare (fraud, churn, disease, anomalies). Covers correct metrics, resampling (SMOTE/undersampling), class weights, threshold tuning, and avoiding the accuracy trap and resampling leakage.