JBrussee/gemma-4-31B-caveman

Model Overview

JBrussee/gemma-4-31B-caveman is a 31 billion parameter Gemma 4 model, specifically fine-tuned to produce highly concise, 'caveman-style' text. This unique output format removes articles, filler words, and pleasantries, favoring fragmented sentences while strictly preserving the byte-exactness of technical elements like code blocks, function names, error strings, and CLI commands. It functions as a direct replacement for google/gemma-4-31B-it but with its distinct linguistic style.

Key Capabilities

• Concise Output: Generates responses in a stripped-down, direct style, ideal for quick comprehension.
• Technical Accuracy: Ensures byte-exact preservation of code, commands, and technical terms within its simplified language.
• Semantic Preservation: Achieves high semantic similarity (91-98%) despite significant text compression.
• Code-Oriented: Excels in tasks like debugging, code review, and refactoring, as demonstrated by its 96-100% code fence exactness.

Training and Performance

The model was fine-tuned using QLoRA NF4 with Unsloth, on a dataset of 1750 train and 193 evaluation pairs. This data was created by transforming content from permissive datasets (e.g., OpenAssistant/oasst2, princeton-nlp/SWE-bench_Verified) into the caveman style using Claude Code and Codex CLI. Evaluation shows strong performance in code preservation and low article density (0.5-2%), with a slight weakness in compression ratio compared to gold-standard caveman pairs.

Use Cases

• Developer Tools: Integrate into IDEs or CLI tools for brief, actionable explanations.
• Technical Documentation: Generate simplified summaries or instructions.
• Code Review: Provide direct, no-nonsense feedback on code.
• Debugging Assistance: Offer concise diagnostic information without verbose explanations.