Gemma 4 E4B fine-tune reaches 94% JSON accuracy

While Gemma 4 E4B is natively strong at JSON, this project demonstrates that targeted fine-tuning is still essential for high-precision document parsing in specialized domains. The fine-tune achieved a significant 19% boost in document type classification accuracy with only 432 training examples and halved hallucinated obligations from 1.25 to 0.59 per document while maintaining 100% valid JSON output. It also highlights critical implementation "gotchas" with Gemma 4's multimodal processor, requiring specific keyword-arg handling to avoid crashes.