The Toolkit

A degree-7 polynomial fitted to a noisy sine wave. With 20 points the training error is 0.050 and the validation error 0.155; with 80, 0.079 and 0.102; with 320, 0.094 and 0.095. Training error rises and validation error falls until they meet near the noise level (0.09). A gap that is still closing says more data will help.

• Curves from a single split are noisy; average several before reading a trend into them.
• Learning curves for different models cross: the best model for small data is often not the best for large (Perlich, Provost & Simonoff 2003).
• A flat validation curve can mean the model is limited, or that the labels are too noisy to learn more from.

rng = np.random.default_rng(3)
f = lambda x: np.sin(3 * x)
x_val = rng.uniform(-1, 1, 2_000); y_val = f(x_val) + rng.normal(0, 0.3, 2_000)
gaps = {}
for n in (20, 80, 320):
    x = rng.uniform(-1, 1, n); y = f(x) + rng.normal(0, 0.3, n)
    c = np.polyfit(x, y, 7)
    train = ((np.polyval(c, x) - y) ** 2).mean(); val = ((np.polyval(c, x_val) - y_val) ** 2).mean()
    gaps[n] = (round(train, 3), round(val, 3))

Assumes import numpy as np and import pandas as pd (and SciPy where imported). Each snippet simulates or makes up its own data, as the comments say.

# Python — learning curves
from sklearn.model_selection import learning_curve
import matplotlib.pyplot as plt

sizes, train_scores, val_scores = learning_curve(
    model, X, y, cv=5, n_jobs=-1,
    train_sizes=np.linspace(0.1, 1.0, 10),
    scoring='accuracy'
)
plt.plot(sizes, train_scores.mean(axis=1), label='Train')
plt.plot(sizes, val_scores.mean(axis=1), label='Val')
plt.legend(); plt.show()