LMSteinshark Small
A decoder-only language model built in PyTorch. The current work emphasizes reproducible checkpoints, mixed data, measurable training runs, instruction tuning, retrieval, and tools.
Read the current update →Independent software and machine-learning projects
Build the thing. Understand it. Write down what broke.
A working notebook of language-model training, chess reinforcement learning, data pipelines, inference systems, and the engineering lessons hidden between the headline results.
Active work
AlphaZero-Inspired Chess
A policy-and-value network paired with batched PUCT search, millions of indexed positions, sparse replay storage, self-play, and arena promotion.
Read the current update →Why document it?
Real implementation details
Architecture, datasets, throughput, state restoration, model serving, action encodings, and mistakes that looked plausible enough to survive for weeks.
Historical record
Older experiments remain available, clearly labeled. The point is not to pretend every first approach was correct; it is to show what changed and why.