Praxis Chess

Small language models that explain why a chess move was a mistake, using only verified engine evidence, with every claim checked by a program.

• Collection: Praxis Chess LoRA v1
• Models: Qwen3.5 2B and 4B LoRA adapters, GGUFs for Ollama / llama.cpp, and the comparison arms
• Dataset: evidence graphs, four input renderings, verified targets, and every published answer
• Reproduce: REPRODUCE.md; re-scoring needs no GPU
• App: Praxis Chess on GitHub, a fully offline chess-improvement system

Only public Lichess data is published.