nislam-mics/ATLAS-Brain-Build-LoRA

ATLAS-NIST-Measure Model Card

Model Description

This is a fine-tuned version of Llama-3.1-8B-Instruct, optimized using Unsloth with 4-bit quantization (LoRA). It serves as the core risk assessment engine for the ATLAS System, designed to classify user actions based on contextual risk levels.

Intended Use

This model is specifically engineered for the Week 6 Control Layer of the ATLAS architecture. It acts as a decision engine that intercepts user actions and outputs a risk classification JSON.

Risk Classification Logic

The model classifies input events into three distinct categories:

1. ROUTINE: Low-risk actions (e.g., login from trusted device, small transactions).
   • Action: Allow request to proceed.
2. ESCALATE: Medium-risk or anomalous actions (e.g., high-value transfers, login from new location).
   • Action: Trigger Human-in-the-Loop review via MCP (Model Context Protocol).
   • Compliance: Aligns with EU AI Act Article 14 requiring human oversight for high-risk AI systems.
3. BLOCK: High-risk or malicious actions (e.g., SQL injection, sanctioned entity transfer).
   • Action: Immediately terminate session (403 Forbidden).

Compliance & Safety

• NIST AI RMF: Implements the 'Measure' function by providing quantifiable risk outputs.
• EU AI Act: Enforces Article 14 (Human Oversight) by flagging ambiguous cases as 'ESCALATE' rather than making autonomous high-stakes decisions.
• Safety: The model has been verified against the 'Alex' persona scenario (redundant payments) to ensure correct blocking behavior.

Usage Instructions

To load this model using Unsloth for the Control Layer:

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "nislam-mics/ATLAS-NIST-Measure",
    model_seq_length = 2048,
    dtype = None,
    load_in_4bit = True,
)
FastLanguageModel.for_inference(model)