Numerical instability in log‑based loss functions

[github-actions]

File: leanpass/nn.py

Functions Tensor.log, cross_entropy_loss and binary_cross_entropy_loss compute np.log of probabilities that can be exactly zero, producing -inf or nan gradients. This is a common source of crashes during training.

Fix: clamp inputs with a small epsilon (e.g., eps = 1e-12) before taking the log, and document the behavior.

Label: enhancement

Filed automatically by ai-issue-scan.