Dwarf-Bench
An LLM benchmark that grades models on obscure dwarf trivia from across fantasy media. A learning exercise in building an eval pipeline from scratch.
An LLM benchmark that grades models on obscure dwarf trivia from across fantasy media. A learning exercise in building an eval pipeline from scratch.
A GPT-2-scale model that speaks English and Klazan, trained by mixing a synthetic Klazan corpus into a nanochat pretraining run. Ablations measure how much dwarf a model can learn before its English starts to suffer.
A homebrew dwarf language inspired by Khazalid, built to be quick for a human to learn and fully machine-generable: phonology, grammar, a 3,000-word lexicon, and a deterministic English to Klazan translator that writes dwarfgpt's training data.
A focus-hours watcher for social platforms: a Firefox extension and a local daemon that block YouTube and X during the workday. A blocked page can be appealed in one sentence, and Claude decides whether it opens for a few minutes.