norecyc/lastbox-gemma4-e2b-sft-v3

VISIONConcurrency Cost:1Model Size:5.1BQuant:BF16Ctx Length:32kTool Calling:SupportedPublished:May 19, 2026License:gemmaArchitecture:Transformer Cold

The norecyc/lastbox-gemma4-e2b-sft-v3 model is a 5.1 billion parameter Gemma 4 E2B variant, fine-tuned by Mateusz Pawelczuk for offline survival assistance. It is optimized for deployment on resource-constrained devices like the Raspberry Pi 5, providing survival Q&A, optical triage, and mesh-radio relay capabilities. This model is specifically designed for low-latency, low-bandwidth communication in offline environments, with responses hard-capped at 150 bytes for LoRa packet compatibility.

Loading preview...

LastBox v3: Offline Survival Assistant

This model, norecyc/lastbox-gemma4-e2b-sft-v3, is a 5.1 billion parameter Gemma 4 E2B variant, fine-tuned by Mateusz Pawelczuk for the Kaggle "Gemma 4 Good" Hackathon 2026. It's engineered for offline survival assistance and designed to run efficiently on an 8 GB Raspberry Pi 5, achieving approximately 700 ms first token latency and 6–7 tok/s sustained throughput on an ARM CPU.

Key Capabilities

  • Survival Q&A: Provides text-only information on first aid, bushcraft, navigation, power, and hazards.
  • Optical Triage: When paired with mmproj-F16.gguf (SigLIP vision), it can describe plants, wounds, and terrain from a Raspberry Pi camera.
  • Mesh-Radio Relay: Replies are hard-capped at 150 bytes UTF-8 to fit single LoRa 868 MHz mesh radio packets, ensuring compatibility with low-bandwidth communication.

Training and Performance

The model was fine-tuned using Unsloth FastModel and TRL SFTTrainer on 1,034 English survival dialogs. It utilizes a gemma-4 chat template (no-thinking) to suppress Chain-of-Thought preambles. While this v3 snapshot has a ~0% tool_emission_rate due to a missing tool-definitions JSON block during evaluation, a subsequent v6 version addresses this. The model demonstrates a response_quality of 0.506 and byte_compliance of 0.48 in its evaluation.

Good for

  • Edge AI applications requiring low-resource deployment.
  • Offline assistance systems in remote or emergency scenarios.
  • Low-bandwidth communication where message size is critical.
  • Prototyping survival-oriented AI on Raspberry Pi platforms.