Gemma 4 LoRa Fine-Tuning

I want to push the Gemma 4 E2B model to its limits for my HA MCP tools by applying LoRa fine-tuning on a purpose-built dataset. All training data must come from manual data entries—no scraping or sensor logs—because I need full control over quality and privacy.

Scope of the dataset
• 500 single-tool examples for every HA MCP tool
• 100 multi-tool workflow examples for each tool group
• Overall target: roughly 15 k – 20 k well-structured prompts and ideal completions

What I expect from you
1. Design an efficient schema that separates single-tool tasks from multi-tool workflows yet still feeds cleanly into the Gemma 4 LoRa pipeline.
2. Build and validate the dataset (CSV/JSONL preferred).
3. Implement LoRa fine-tuning on the Gemma 4 E2B base, iterating until the model converges.
4. Track and report Accuracy, Precision, and Recall at each epoch; include plots plus a concise summary of the best checkpoint.
5. Deliver: curated dataset, training scripts (Python; Hugging Face or compatible), final LoRa weights, and an evaluation report.

If this aligns with your skill set in NLP, dataset engineering, and model optimisation, let’s get started.