[ARCHIVED] This repository is kept for historical reference. It represents an early experiment (~1394 parameters, NumPy/CuPy). The current generation neural network project is nn-v4 — a 1.5B-parameter GPT-style transformer.

A primitive character-level language model using NumPy with GPU acceleration (CuPy). Part of a learning progression: nn-v1 → nn-v2 → nn-v4.

• ~1000 parameters: Ultra-lightweight neural network
• GPU acceleration: Uses CuPy for NVIDIA GPU support (falls back to NumPy on CPU)
• Character-level: Learns patterns at character granularity
• Simple architecture: Embedding → Average Pooling → Hidden Layer (ReLU) → Output

�ash pip install -r requirements.txt python train.py # train + save model python generate.py # interactive text generation "@

= @"

[ARCHIVED] This repository is kept for historical reference. It represents an intermediate experiment (~6.8K parameters, PyTorch bigram model). The current generation neural network project is nn-v4 — a 1.5B-parameter GPT-style transformer.

A minimal character-level bigram language model using PyTorch. Part of a learning progression: nn-v1 → nn-v2 → nn-v4.

• model.py — BigramModel class
• rain.py — Training script (creates �igram_model.pth)
• generate.py — Interactive text generation
• data.txt — Shakespeare training text

�ash pip install -r requirements.txt python train.py # train for 5000 iterations python generate.py # interactive generation "@

= @"

[ARCHIVED] This is a legacy copy of Andrej Karpathy's nanoGPT used for early experiments. The current generation neural network project is nn-v4 — a 1.5B-parameter GPT-style transformer built from scratch.

An early experiment following Andrej Karpathy's nanoGPT to understand GPT-style transformer training.

Key learnings from this stage fed into the design of nn-v4.