# GloVe Code Walkthrough (NumPy)

This page walks through the full GloVe training pipeline from raw text to embeddings.

## Code sections

1. Corpus preprocessing and vocabulary build
2. Co-occurrence matrix creation with distance weighting
3. Parameter initialization (`W`, `W_tilde`, biases)
4. Weighted least squares objective
5. Manual gradient computation
6. AdaGrad updates
7. Metrics export for visualization

## Practical notes

- Training iterates only over non-zero co-occurrence entries.
- AdaGrad stabilizes optimization for Zipfian frequency distributions.
- Final embeddings are typically formed as `W + W_tilde`.