README.md

metadata

title: MindForge AI
emoji: 🧠
colorFrom: blue
colorTo: purple
sdk: gradio
sdk_version: 4.44.0
python_version: '3.11'
app_file: app.py
pinned: false
license: mit
short_description: MindForge AI — AMD hackathon distress & escalation demo

MindForge AI

MindForge AI is a human-in-the-loop distress detection and conversational escalation engine built for the AMD Developer Hackathon. It fine-tunes Qwen for one specific task: converting mental-health check-ins and care notes into structured risk-review outputs and safe escalation actions.

What it does

• Classifies distress level from 0 to 3
• Selects the safest next action
• Extracts risk and protective signals
• Flags medication, adherence, sleep, and mood concerns
• Routes to pre-vetted response templates
• Produces a care-team summary
• Keeps a human in the loop

Safety boundary

MindForge AI does not diagnose, treat, or replace licensed care. It organizes distress signals and recommends escalation pathways for human review.

Fine-tuning methodology

• Base model: Qwen Instruct
• Method: LoRA supervised fine-tuning
• Dataset: synthetic safety-labeled mental-health check-ins
• Compute target: AMD Developer Cloud, ROCm, AMD Instinct MI300X
• Evaluation: JSON validity, distress accuracy, action F1, crisis recall, unsafe-overreach rate

Built with

• Gradio
• Hugging Face Spaces
• Qwen
• AMD Developer Cloud
• ROCm
• PyTorch
• LoRA / PEFT

Space configuration

Optional environment variables (for live base-model comparison only; the demo runs without them using the rules classifier):

Variable	Purpose
USE_MODEL	true / false — attempt optional transformers inference
BASE_MODEL_ID	Hugging Face model id for optional base generation
ADAPTER_MODEL_ID	Optional PEFT adapter id
HF_TOKEN	Private model access token if needed

Crises routing and structured JSON outputs always use the deterministic rules engine + templates so the Space stays safe and reliable even when weights are not loaded.