# TinyBuddy-30M

A 30 million parameter GPT-style transformer trained on TinyStories.

## Architecture
- 6 layers, 8 attention heads, 256 embedding dim
- 50,000 vocabulary size (untied weights)
- 512 context length (trained on 128 for speed)

## Training
- Dataset: TinyStories (5,000 stories)
- Steps: 1,500
- Hardware: CPU only
- Loss: ~5.5 (coherent but not good)

## What It Can Do
- Generate 2-3 word fragments that resemble story patterns
- Sometimes repeat words from the prompt
- Produce gibberish that's trying to be English

## What It Cannot Do
- Tell a coherent story
- Answer questions
- Anything useful

## Why It Exists
To demonstrate that even a tiny transformer learns *patterns*, not rules.
This is a real AI, just a very small one.