📦 LASTBOX // CHAT
Offline survival assistant on a Raspberry Pi 5, fine-tuned Gemma 4 E2B. This Space runs the v6 SFT-warmup checkpoint (72 % tool emission, agentic_score 0.608) on ZeroGPU.
📦 GitHub · 🌐 Project site
Heads-up: cold-start ~30 s. Warm responses 5–20 s. ZeroGPU has a shared free-tier GPU pool — if you see "Model generation failed — ZeroGPU may be queued", wait 20-30 s and resend. The LoRa Radio tab runs two model turns per request (tool call → result → final reply), so it burns quota faster than Free/RAG.
Terse replies under the 150-byte UTF-8 cap so the answer fits in one LoRa packet. The model emits <tool_call> blocks when it would dispatch a tool on the real device.
Long-form survival Q&A. Crisis-mode system prompt — actionable answers, no 'consult a professional' hedging.
Retrieves top-4 passages from the 4 074-passage offline corpus (train_v2 + Wikipedia survival/first-aid) and cites source IDs.
About this Space. This is a public chat demo of the LastBox v6 checkpoint. The real device is a Pelican-cased Raspberry Pi 5 with a LoRa 868 MHz HAT, a camera, and a battery; everything runs offline. Numbers on the project site are measured against the actual on-device deployment.
License. Code/LoRA Apache 2.0 · weights subject to Gemma Terms of Use.