# Inside GloVe: Matrix, Bias, and Factorization Analysis

This section inspects the trained GloVe model with real run artifacts.

## Diagnostics included

- Co-occurrence matrix sparsity and top pair structure
- Factorization quality over epochs (reconstruction metrics)
- `W` vs `W_tilde` cosine alignment
- Bias-frequency correlation checks
- PMI vs dot-product comparison
- Neighbor and analogy evolution during training

## Why these internals matter

- They verify the model is learning meaningful structure, not only lowering loss.
- They confirm theoretical claims (frequency in biases, semantics in vectors).
- They make it easier to debug hyperparameters such as window size, alpha, and learning rate.