te-nims-e4b-stage9-gguf

TE-NIMS E4B Stage 9 — Q4_K_M GGUF

Fine-tuned Gemma 4 E4B for NIMS / ICS decision support in civilian emergency management. Trained through 9 warm-start SFT stages on curated FEMA doctrine, ICS forms, and after-action reports.

• Base: google/gemma-4-e4b-it (dense 4B edge model)
• Adapter lineage: tmancino/te-nims-e4b-stage9
• Quantization: Q4_K_M (~5GB, runs on 8GB VRAM)
• Training: 9-stage warm-start SFT chain (Stage 9 is the production checkpoint)
• ODA eval score: 0.916 on the 52-case TE NIMS Bench

This repository is the deployable text inference artifact. It is the model downloaded by the TE NIMS demo on first boot and wrapped by Ollama as the local runtime severian-ollama.

Use cases

• Incident Commander decision support
• ICS-201 / ICS-202 / ICS-213 form generation
• NIMS doctrine grounding for first-responder questions
• Unified Command planning and mass-casualty triage

Quickstart

Ollama

ollama run hf.co/tmancino/te-nims-e4b-stage9-gguf

Docker (full FOB demo UI)

git clone https://github.com/TerminusEstAI/te-nims-demo
cd te-nims-demo && docker compose up
open http://localhost:8765

llama.cpp

huggingface-cli download tmancino/te-nims-e4b-stage9-gguf nims-e4b-stage9-q4_k_m.gguf
./llama-server -m nims-e4b-stage9-q4_k_m.gguf -c 8192

Live demo

Full Incident Commander UI: https://demo.terminusest.ai

Source: https://github.com/TerminusEstAI/te-nims-demo

Deployment notes

• Demo runtime name: severian-ollama
• Intended runtime: Ollama or llama.cpp-compatible GGUF serving
• The demo's multimodal severian-vision sidecar is a separate deployment path with separate vision artifacts

Citation

Built by Terminus Est AI for the Kaggle Gemma 4 Good Hackathon, May 2026.

License

CC BY 4.0 — https://creativecommons.org/licenses/by/4.0/

Copyright (c) 2026 Terminus Est AI, Inc.

Base model google/gemma-4-e4b-it retains its original Apache 2.0 license and Gemma Terms of Use.