writing

A Stanford CS224N project exploring whether a council of fine-tuned GPT-2 models can improve structured text generation. The strongest gains show up in sonnet writing, where debate-guided generation substantially improves metrical adherence over single-model and simpler ensemble baselines.

LLMs default to bulleted lists optimized for visual reading, which sound unnatural in voice interfaces. We fine-tune Llama 3.2 1B with GRPO using synthetic preference pairs to produce TTS-friendly prose, achieving a 66.7% reduction in list usage.

There's no feeling like it

This was a test-case for development but left in to display markdown capabilities lol

A deep dive into the mathematics behind Support Vector Machines — illustrated with geometric visualizations, graphs, and worked equations — demonstrated through two real-world classification problems: Australian weather prediction and video game rating prediction.