To Train

This fine-tunes NVIDIA Nemotron 3.5 Bolt E4B (GRPO post-trained Gemma 4 E4B) on the Home-Assistant-Requests-V2 dataset. Nemotron is already optimized for tool use and instruction following — a much better starting point than base Gemma 4 E4B-it for HA intent recognition.

Download model into cache

uv run huggingface-cli download nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-BF16 --local-dir ./models/nemotron-e4b

Run training

# Test run (1000 rows)
uv run python train.py --test

# Full run (209k rows)
uv run python train.py

Export GGUF (standalone, if training already completed)

uv run python export_gguf.py

The exported Q4_K_M GGUF is optimized for the Jetson Orin Nano 8GB:

• mmap keeps PLE embeddings evicted under memory pressure (~3GB active)
• Deploy to /home/lalobied/models/ha-gemma4-e4b.gguf on the Jetson