cavegemma
A LoRA fine-tune of Gemma that answers in a terse, compressed style to cut output tokens, while keeping code blocks and meaning intact, evaluated on held-out prompt pairs.
cavegemma is a LoRA fine-tune of Gemma 4 31B trained to answer in a heavily compressed, terse style, without a system prompt or toggle to turn it on. Across 193 held-out pairs it produces 27 percent fewer output tokens on average, while semantic similarity to the verbose original stays at 0.91 to 0.98 cosine and fenced code blocks come back byte-exact 96 to 100 percent of the time.
The reduction is uneven: chat and question-answering responses shrink the most, while long debugging and refactoring walkthroughs barely change, because the training data filter allowed rewrites up to the full length of the source and the model learned to sit near that limit.
All reported numbers are reproducible by running the project's own evaluation script against its seed pairs.