A fine-tuned Gemma 4 model for local, privacy-preserving AI safety coaching for minors. It classifies AI-chat prompts into coaching areas such as Privacy, Safety, Wellbeing, and Learning, then returns JSON for a local report without storing raw prompts.

tools thinking

ollama run bzwear/gemma-prism:v2-q4

curl http://localhost:11434/api/chat \
  -d '{
    "model": "bzwear/gemma-prism:v2-q4",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from ollama import chat

response = chat(
    model='bzwear/gemma-prism:v2-q4',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'bzwear/gemma-prism:v2-q4',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

Details

Updated yesterday

yesterday

0ff04c2706fb · 3.4GB ·

model

archgemma4

parameters4.65B

quantizationQ4_K_M

3.4GB

system

You are Gemma Prism Compact, a local AI safety classifier. Return compact JSON for privacy-preservin

111B

params

{ "num_ctx": 2048, "stop": [ "<turn|>" ], "temperature": 0.2 }

64B

Readme

Gemma Prism

Fine-tuned Gemma 4 model for Gemma Prism, a local-first AI safety coach for minors using AI chat tools.

Use

ollama pull bzwear/gemma-prism:v2-q4
ollama run --think=false --format json bzwear/gemma-prism:v2-q4

This model is used by the Gemma Prism local FastAPI runtime:

https://github.com/bzwear/gemma-prism

Intended Use Gemma Prism converts local AI-chat signals into compact, privacy-preserving safety guidance. It is designed for youth-facing coaching reports, not surveillance or raw prompt monitoring.

Output Contract The app expects compact JSON with category, risk_level, confidence, summary, nudge, and rewrite fields. The local server validates, normalizes, redacts, and applies deterministic safety policy before rendering reports.