AEGIS is a locally deployable Gemma 4 E4B instruction-tuned model package for scam-defense analysis. It is tuned for one job: analyze a phone-call transcript for epistemic manipulation and return structured JSON describing how a scammer is trying to distort the target's decision-making process.

This repository is the Ollama / GGUF release package for:

Local inference with Ollama
Offline demos and hackathon judging
Safety and trust evaluation on scam-call transcripts

The package is designed for the GOVINDFROM/aegis-gemma-4-e4b-it-scam-defense-ollama-gguf-v1 Hugging Face repo.

What This Model Does

Given a conversation transcript, AEGIS scores six manipulation vectors:

belief_installation
verification_suppression
urgency_fabrication
authority_hijacking
emotional_flooding
exit_path_closure

It returns:

Per-vector scores from 0-100
integrity_score
risk_level
A plain-language explanation
A recommended next action
Evidence snippets per vector

Files

gemma-4-e4b-it.Q4_K_M.gguf: quantized GGUF model for local inference
Modelfile: Ollama model definition with the AEGIS system behavior and prompt template
README.md: usage and model card

Base Model

Base model: google/gemma-4-E4B-it
Deployment format: GGUF
Quantization: Q4_K_M

Training Summary

This release is derived from a fine-tuned Gemma 4 E4B instruction model adapted for scam-defense transcript analysis.

Training method: LoRA
LoRA rank: 16
Learning rate: 2e-4
Epochs: 3
Train size: 396
Eval size: 44
Max sequence length: 2048
Final reported training loss: 0.4065
Training hardware: NVIDIA A100-SXM4-80GB

Recommended Use

Use this model to:

Analyze scam and non-scam phone-call transcripts
Demonstrate local-first AI safety tooling
Power an Ollama-based Streamlit, Gradio, or desktop app
Support structured risk scoring for research demos

Do not use this model as:

A legal authority
A replacement for emergency services
A sole decision-maker in high-stakes financial or criminal matters

Download And Run With Ollama

Option 1: Download from Hugging Face, then create the Ollama model

Download the repository:

bash

git lfs install
git clone https://huggingface.co/GOVINDFROM/aegis-gemma-4-e4b-it-scam-defense-ollama-gguf-v1
cd aegis-gemma-4-e4b-it-scam-defense-ollama-gguf-v1

Create the local Ollama model:

ollama create aegis -f Modelfile

Run it:

ollama run aegis

Option 2: Download with the Hugging Face CLI

bash

huggingface-cli download GOVINDFROM/aegis-gemma-4-e4b-it-scam-defense-ollama-gguf-v1 --local-dir .
ollama create aegis -f Modelfile
ollama run aegis

Example Prompt

Paste a transcript like this into Ollama:

text

Analyze this phone call for epistemic manipulation:

Caller: Hello, this is Agent Williams from the IRS Criminal Investigation Division.
We found major discrepancies in your tax filings.
A federal arrest warrant has been issued in your name.
You must pay immediately via gift cards to stop enforcement.
Do not hang up or contact anyone else.

Expected behavior:

High authority_hijacking
High urgency_fabrication
High verification_suppression
risk_level near CRITICAL
Very low integrity_score

Example Output Shape

The model is configured to emit JSON in this structure:

json

{
  "manipulation_vectors": {
    "belief_installation": 95,
    "verification_suppression": 90,
    "urgency_fabrication": 100,
    "authority_hijacking": 100,
    "emotional_flooding": 85,
    "exit_path_closure": 95
  },
  "integrity_score": 5,
  "risk_level": "CRITICAL",
  "explanation": "The caller is using false authority, panic, and isolation tactics to block independent verification.",
  "recommended_action": "Terminate the call and verify the claim through an official number you find independently.",
  "evidence": {
    "belief_installation": "False claim of tax discrepancies and an arrest warrant.",
    "verification_suppression": "Instructs the target not to contact anyone else.",
    "urgency_fabrication": "Immediate payment demanded.",
    "authority_hijacking": "Claims to be from the IRS Criminal Investigation Division.",
    "emotional_flooding": "Threat of arrest creates fear.",
    "exit_path_closure": "Explicit instruction not to hang up."
  }
}

Hardware Notes

Runs locally with Ollama
Suitable for laptops / desktops
Quantized package size is about 5.3 GB
GPU helps, but CPU inference is possible with higher latency

Limitations

This model analyzes transcript text, not verified ground truth
Scam tactics evolve and can differ by region and language
Structured JSON output may occasionally need retry logic in downstream apps
Scores should support human judgment, not replace it

Safety Notes

This model is intended to help users recognize manipulative scam behavior. It should be deployed with:

Clear user-facing disclaimers
Human override
Independent verification guidance
Privacy-preserving local inference where possible

Citation

If you use this model, cite the project repository and hackathon submission:

bibtex

@misc{aegis_gemma4_2026,
  title={AEGIS: Adaptive Epistemic Guard for Intelligent Scam Defense},
  author={GOVINDFROM},
  year={2026},
  howpublished={Hugging Face model repository}
}