Train

Model Training & Evaluation

Templates and reference for training, evaluating, and persisting ML models.

Data splitting

Standard (>10k rows)

from sklearn.model_selection import train_test_split

X_train, X_temp, y_train, y_temp = train_test_split(X, y, test_size=0.3, random_state=42, stratify=y)
X_val, X_test, y_val, y_test = train_test_split(X_temp, y_temp, test_size=0.5, random_state=42, stratify=y_temp)

Cross-validation (<10k rows)

from sklearn.model_selection import StratifiedKFold, cross_val_score
cv = StratifiedKFold(n_splits=5, shuffle=True, random_state=42)
scores = cross_val_score(model, X, y, cv=cv, scoring='accuracy')
print(f"CV: {scores.mean():.4f} +/- {scores.std():.4f}")

Time series

from sklearn.model_selection import TimeSeriesSplit
tscv = TimeSeriesSplit(n_splits=5)

Model selection guide

Classification

Regression

Decision rule: start simple

Linear/Naive Bayes → SVM/KNN → Tree ensembles → Neural Net
     ↑                                                 ↑
 interpretable                                   highest capacity

Always record a linear baseline as exp000 before trying complex models.

Templates

sklearn pipeline

from sklearn.pipeline import Pipeline
from sklearn.preprocessing import StandardScaler
from sklearn.ensemble import RandomForestClassifier
from sklearn.metrics import classification_report
import joblib

pipe = Pipeline([
    ('scaler', StandardScaler()),
    ('model', RandomForestClassifier(n_estimators=100, random_state=42, n_jobs=-1))
])
pipe.fit(X_train, y_train)
y_pred = pipe.predict(X_val)
print(classification_report(y_val, y_pred))
joblib.dump(pipe, 'model.joblib')

XGBoost

import xgboost as xgb

dtrain = xgb.DMatrix(X_train, label=y_train)
dval = xgb.DMatrix(X_val, label=y_val)
params = {
    'objective': 'binary:logistic', 'eval_metric': 'logloss',
    'max_depth': 6, 'learning_rate': 0.1, 'subsample': 0.8, 'seed': 42,
}
model = xgb.train(params, dtrain, num_boost_round=1000,
    evals=[(dtrain, 'train'), (dval, 'val')], early_stopping_rounds=50, verbose_eval=100)
model.save_model('model.xgb')

PyTorch

import torch
import torch.nn as nn

class MLP(nn.Module):
    def __init__(self, input_dim, hidden=128, output=1):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(input_dim, hidden), nn.ReLU(), nn.Dropout(0.2),
            nn.Linear(hidden, hidden), nn.ReLU(), nn.Dropout(0.2),
            nn.Linear(hidden, output),
        )
    def forward(self, x): return self.net(x)

model = MLP(X_train.shape[1])
opt = torch.optim.AdamW(model.parameters(), lr=1e-3, weight_decay=1e-4)
loss_fn = nn.BCEWithLogitsLoss()

for epoch in range(100):
    model.train()
    for bx, by in loader:
        opt.zero_grad()
        loss_fn(model(bx).squeeze(), by).backward()
        opt.step()
torch.save(model.state_dict(), 'model.pt')

Naive Bayes

from sklearn.naive_bayes import GaussianNB, MultinomialNB, BernoulliNB
from sklearn.metrics import classification_report
import joblib

# GaussianNB — continuous features (assumes Gaussian distribution per class)
model = GaussianNB()
model.fit(X_train, y_train)
print(classification_report(y_val, model.predict(X_val)))
joblib.dump(model, 'model_nb.joblib')

# MultinomialNB — count/frequency features (e.g. TF-IDF, bag-of-words)
# from sklearn.naive_bayes import MultinomialNB
# model = MultinomialNB(alpha=1.0)  # alpha = Laplace smoothing

# BernoulliNB — binary features (word presence/absence)
# from sklearn.naive_bayes import BernoulliNB
# model = BernoulliNB(alpha=1.0)

# Tuning: var_smoothing (GaussianNB), alpha (Multinomial/Bernoulli)
# When to use: text classification, spam detection, small data, fast baseline

K-Nearest Neighbors (KNN)

from sklearn.neighbors import KNeighborsClassifier, KNeighborsRegressor
from sklearn.preprocessing import StandardScaler
from sklearn.pipeline import Pipeline
import joblib

# Classification
pipe = Pipeline([
    ('scaler', StandardScaler()),   # KNN is distance-based — scaling is mandatory
    ('model', KNeighborsClassifier(
        n_neighbors=5,
        weights='distance',         # 'uniform' or 'distance' (closer = more weight)
        metric='minkowski',         # Euclidean when p=2, Manhattan when p=1
        n_jobs=-1,
    )),
])
pipe.fit(X_train, y_train)
joblib.dump(pipe, 'model_knn.joblib')

# Regression
# pipe = Pipeline([('scaler', StandardScaler()), ('model', KNeighborsRegressor(n_neighbors=5))])

# Tuning: n_neighbors (odd to avoid ties), weights, metric
# Weakness: O(n) prediction time — slow at inference on large datasets
# When to use: small/medium data, non-linear boundaries, anomaly detection

Discriminant Analysis (LDA / QDA)

from sklearn.discriminant_analysis import LinearDiscriminantAnalysis, QuadraticDiscriminantAnalysis
from sklearn.metrics import classification_report
import joblib

# LDA — assumes equal covariance across classes; also works as dimensionality reduction
lda = LinearDiscriminantAnalysis(
    solver='svd',          # 'svd' (default), 'lsqr', 'eigen'
    n_components=None,     # reduce to min(n_classes-1, n_features) components
    store_covariance=False,
)
lda.fit(X_train, y_train)
print(classification_report(y_val, lda.predict(X_val)))
joblib.dump(lda, 'model_lda.joblib')

# For dimensionality reduction (supervised):
# X_reduced = lda.transform(X_train)  # reduces to n_classes-1 dimensions

# QDA — allows different covariance per class; more flexible but needs more data
# qda = QuadraticDiscriminantAnalysis(reg_param=0.0)  # reg_param adds regularization

# When to use: Gaussian class distributions, interpretable decision boundary,
#              dimensionality reduction to n_classes-1, well-separated classes

Support Vector Machine (SVM / SVR)

from sklearn.svm import SVC, SVR, LinearSVC
from sklearn.preprocessing import StandardScaler
from sklearn.pipeline import Pipeline
import joblib

# Classification — RBF kernel (best general-purpose choice)
pipe = Pipeline([
    ('scaler', StandardScaler()),   # SVM is distance-based — scaling is mandatory
    ('model', SVC(
        C=1.0,              # Regularization: high C = low bias/high variance
        kernel='rbf',       # 'linear', 'poly', 'rbf', 'sigmoid'
        gamma='scale',      # 'scale' = 1/(n_features*X.var()), 'auto' = 1/n_features
        probability=True,   # enables predict_proba (slower fitting)
        random_state=42,
        class_weight='balanced',  # handles class imbalance
    )),
])
pipe.fit(X_train, y_train)
print(pipe.predict_proba(X_val)[:5])
joblib.dump(pipe, 'model_svm.joblib')

# For large datasets (>10k): use LinearSVC (much faster, linear kernel only)
# pipe = Pipeline([('scaler', StandardScaler()), ('model', LinearSVC(C=1.0, max_iter=2000))])

# Regression — SVR
# pipe = Pipeline([('scaler', StandardScaler()), ('model', SVR(C=1.0, epsilon=0.1, kernel='rbf'))])

# Tuning priority: C first, then gamma (for RBF), then kernel
# When to use: small/medium data (<50k), high-dimensional (text, images), clear margin of separation

Decision Tree

from sklearn.tree import DecisionTreeClassifier, DecisionTreeRegressor, export_text
from sklearn.metrics import classification_report
import joblib

# Classification
tree = DecisionTreeClassifier(
    max_depth=5,            # constrain depth to prevent overfitting
    min_samples_split=20,   # min samples to split a node
    min_samples_leaf=10,    # min samples in a leaf
    criterion='gini',       # 'gini' or 'entropy'
    class_weight='balanced',
    random_state=42,
)
tree.fit(X_train, y_train)
print(classification_report(y_val, tree.predict(X_val)))

# Print interpretable rules
rules = export_text(tree, feature_names=list(X_train.columns))
print(rules[:2000])   # first 2000 chars of rule set
joblib.dump(tree, 'model_tree.joblib')

# Regression
# tree = DecisionTreeRegressor(max_depth=5, min_samples_leaf=10, random_state=42)

# Tuning priority: max_depth → min_samples_leaf → criterion
# Weakness: high variance (small data changes flip the tree) — prefer ensemble unless interpretability required
# When to use: interpretability is mandatory, rule extraction, feature selection proxy

Generalized Linear Model (GLM)

import statsmodels.api as sm
import numpy as np

# Poisson GLM — count data (events per unit time/area)
X_train_sm = sm.add_constant(X_train)   # statsmodels needs explicit intercept
glm_poisson = sm.GLM(
    y_train,
    X_train_sm,
    family=sm.families.Poisson(link=sm.families.links.Log()),
)
result = glm_poisson.fit()
print(result.summary())
y_pred = result.predict(sm.add_constant(X_val))

# Gamma GLM — positive continuous, right-skewed (insurance claims, durations)
# glm_gamma = sm.GLM(y_train, X_train_sm, family=sm.families.Gamma(link=sm.families.links.Log()))

# Tweedie GLM — flexible family (p=0: Normal, p=1: Poisson, p=2: Gamma, 1<p<2: compound)
# glm_tweedie = sm.GLM(y_train, X_train_sm, family=sm.families.Tweedie(var_power=1.5))

# Negative Binomial — overdispersed count data (variance > mean)
# glm_nb = sm.GLM(y_train, X_train_sm, family=sm.families.NegativeBinomial())

# Save coefficients
import json
coefs = dict(zip(['intercept'] + list(X_train.columns), result.params))
with open('model_glm_coefs.json', 'w') as f:
    json.dump(coefs, f, indent=2)

# When to use: count data, rate data, insurance/actuarial, non-Gaussian errors,
#              heteroscedastic residuals, log/logit link needed for interpretability

Gaussian Process (GP)

from sklearn.gaussian_process import GaussianProcessClassifier, GaussianProcessRegressor
from sklearn.gaussian_process.kernels import RBF, Matern, WhiteKernel, ConstantKernel
from sklearn.preprocessing import StandardScaler
from sklearn.pipeline import Pipeline
import joblib
import numpy as np

# Regression — returns mean prediction AND uncertainty (std dev)
kernel = ConstantKernel(1.0) * Matern(length_scale=1.0, nu=2.5) + WhiteKernel(noise_level=0.1)
gpr = GaussianProcessRegressor(
    kernel=kernel,
    alpha=1e-6,          # numerical stability
    normalize_y=True,    # subtract mean of y_train
    n_restarts_optimizer=5,  # restarts to find global kernel hyperparams
    random_state=42,
)
gpr.fit(X_train, y_train)
y_pred, y_std = gpr.predict(X_val, return_std=True)
print(f"Val RMSE: {np.sqrt(np.mean((y_pred - y_val)**2)):.4f}")
print(f"Mean uncertainty (std): {y_std.mean():.4f}")
joblib.dump(gpr, 'model_gp.joblib')

# Classification — probabilistic predictions
# kernel = ConstantKernel(1.0) * RBF(length_scale=1.0)
# gpc = GaussianProcessClassifier(kernel=kernel, n_restarts_optimizer=5, random_state=42)
# gpc.fit(X_train, y_train)
# probs = gpc.predict_proba(X_val)

# Weakness: O(n³) training, O(n²) memory — not viable above ~5k samples
# When to use: uncertainty quantification is required, small data (<5k),
#              spatial/temporal data (use Matern kernel), active learning, Bayesian optimization

Metrics reference

Evaluation report format

=== Training Report ===
Task: Binary Classification
Model: XGBoost (1000 rounds, early stopped at 347)
Split: 35k train / 7.5k val / 7.5k test

Val:  Accuracy=0.8634, F1=0.8521, AUC=0.9234
Test: Accuracy=0.8601, F1=0.8489

Top features: feature_a (0.234), feature_b (0.189), feature_c (0.156)
Saved: model.xgb, metrics.json

Rules

• Never evaluate on training data
• Set random seeds everywhere
• Use early stopping for iterative models
• Save both model and preprocessing (use Pipeline or save scaler separately)
• Stratify splits for classification
• Only touch test set ONCE for final evaluation

Experiment Tracking & Iteration

Setup

echo -e "id\tmetric\tval_score\ttest_score\tmemory_mb\tstatus\tdescription" > results.tsv
echo -e "exp000\taccuracy\t0.8523\t0.8401\t4096\tKEEP\tbaseline" >> results.tsv

Results format (TSV)

id        metric    val_score  test_score  memory_mb  status   description
exp000    accuracy  0.8523     0.8401      4096       KEEP     baseline
exp001    accuracy  0.8612     0.8498      4096       KEEP     lr=0.001
exp002    accuracy  0.8590     -           4096       DISCARD  lr=0.003 (overfit)
exp003    accuracy  0.0000     -           0          CRASH    lr=0.01 (diverged)
exp004    accuracy  0.8634     0.8521      4352       KEEP     dropout=0.1

Status: KEEP (improved), DISCARD (same or worse), CRASH (error/OOM/NaN)

Experiment cycle

1. Hypothesize (what change, why it might help)
2. Modify (one variable at a time)
3. Run (fixed budget: time or epochs)
4. Record (append to results.tsv)
5. Decide: KEEP or DISCARD
6. Repeat

What to try (priority order)

High impact (try first)

1. Learning rate (3x and 0.3x current)
2. Model capacity (layers, hidden size)
3. Batch size (double or halve)
4. Regularization (dropout, weight decay)

Medium impact 5. Optimizer (Adam → AdamW → SGD+momentum) 6. LR schedule (cosine, warmup, step decay) 7. Data augmentation 8. Feature selection

Low impact (try last) 9. Activation functions 10. Normalization layers 11. Initialization schemes 12. Gradient clipping

Search strategies

Grid (small spaces)

from itertools import product
params = {'lr': [1e-4, 3e-4, 1e-3], 'dropout': [0.0, 0.1, 0.3]}
for combo in product(*params.values()):
    config = dict(zip(params.keys(), combo))

Random (large spaces)

import random
def sample():
    return {
        'lr': 10 ** random.uniform(-5, -2),
        'dropout': random.uniform(0, 0.5),
        'hidden': random.choice([64, 128, 256, 512]),
    }

Informed (after several runs)

Analyze results.tsv: which LR range works? Did more capacity help? Narrow search.

Analyze results

uv run ${CLAUDE_SKILL_DIR}/scripts/analyze_results.py results.tsv

Or inline:

import pandas as pd
r = pd.read_csv("results.tsv", sep="\t")
kept = r[r.status == "KEEP"]
print(f"Total: {len(r)}, Kept: {len(kept)}, Best: {kept.val_score.max():.6f}")
print(kept.nlargest(5, 'val_score')[['id', 'val_score', 'description']])

Experiment rules

• ONE variable per experiment
• Validation set for decisions, test set only at the end
• Track memory — OOM means not viable
• Fixed random seeds
• Log everything (stdout/stderr to run.log)
• Commit code before each experiment

Code style

Follow the ML code style conventions in references/ml-code-style.md when writing or reviewing training code. Key rules:

• Annotate tensor shapes in docstrings using standard dimension symbols (B, T, D, C, etc.)
• Two-line file header: purpose + author
• Section dividers (# ── Name ──) for files > 200 lines
• No defensive validation inside trusted internals — only at system boundaries
• Document what a variable IS, not what the operation does