Safe4All core model powers every AI experience across the Safe4All platform—delivering adaptive, respectful, and safe assistance to individuals with diverse abilities, cognitive styles, and communication preferences.

🛡️ Safe4All – Main Model

The Safe4All core model powers every AI experience across the Safe4All platform—delivering adaptive, respectful, and safe assistance to individuals with diverse abilities, cognitive styles, and communication preferences.

Field	Value
Base Model	DeepSeek‑R1 7B (Q4_K_M) (fine‑tuned)
Context Length	128 k tokens
License	Apache‑2.0 (inherits base)
System Prompt	Embedded in Modelfile – see below

🎯 Mission

Support independence and self‑expression across physical, cognitive, and emotional needs.
Provide adaptive, safe, and respectful interactions aligned with user capabilities.
Act as a reliable helper, companion, and guide for daily living, learning, communication, and exploration.

🧠 General Behavior Highlights

Kind, clear, and affirming language.
Step‑by‑step guidance with options for BASIC (plain) and DETAIL (expanded) communication modes.
Safety‑first: no medical/legal advice, proactive risk flagging, and gentle escalation.

🛠️ Supported Use‑Case Categories

Daily Living Assistance – hygiene, dressing, cooking, visual schedules.
Emotional Regulation – grounding techniques, check‑ins, affirmations.
Learning & Cognitive Tasks – reading, math, logic tutoring with scaffolding.
Social Interaction Practice – conversation rehearsal, social rules coaching.
Communication Aids – AAC‑friendly formats, symbol support.
Caregiver Tools – progress summaries, adaptation strategies.

🔐 Safety & Boundaries

Never provide medical dosages, diagnoses, or dangerous instructions.
Redirect users to trusted resources or caregivers when appropriate.
Escalate via platform APIs (GuardianBot) if self‑harm or imminent danger signals appear.

🗄️ Memory & Personalization

When memory is enabled, the model can remember user preferences (name, pronouns, tone, accommodations, calming routines) to tailor future interactions. If memory is disabled, no personal data is retained beyond the session.

🤖 Multi‑Agent Integration

Safe4All can delegate domain‑specific tasks to specialized sub‑models:

Agent	Role
CareAgent	Daily routines, wellness, reminders
LearnAgent	Education, cognitive tasks
SocialAgent	Interaction coaching & practice
AssistAgent	Interface navigation, accessibility
GuardianBot	Safety checks & escalation logic

🚀 Quick Start

# Pull the model
ollama pull safe4all-main        # name can vary in your registry

# Run interactively
ollama run safe4all-main

Example Session

User  : “Set up a simple visual schedule for my morning routine.”
Safe4All: Sure! Let’s start with BASIC mode…

📝 System Instruction Prompt (Embedded)

# 🛡️ Safe4All System Instruction Prompt
You are the AI assistant within the Safe4All ecosystem...
(Full prompt continues in Modelfile)

Tip: Keep the full prompt inside the Modelfile under the SYSTEM block to ensure consistent behavior across deployments.

🏗️ Build Locally

# Folder structure
models/
  safe4all-main/
    Modelfile
    README.md     # ← this file

# Build
ollama create safe4all-main -f models/safe4all-main/Modelfile

🛡️ Policy

This model adheres to NeuroEquality’s Responsible AI Policy. All interactions are logged and audited according to Safe4All governance standards.