🧠 MindSpark-1.0

A tiny, ethical AI model that understands child emotions, mindset, and risk signals from short journal entries — 100% offline, zero PII, and Chromebook-ready.

🔒 Built for schools, counselors, and parents who care about privacy, safety, and emotional insight — without surveillance.

---

📦 Model Details

• Architecture: Google BERT-Tiny (4-layer, 256 hidden) + custom multi-task heads
• Tasks:
  • Emotion classification: 7 classes (happy, sad, anxious, angry, lonely, scared, confused)
  • Mindset detection: 6 types (growth, fixed, resilient, helpless, optimistic, pessimistic)
  • Risk flag: Binary (flags phrases like "want to disappear", "nobody would care")
• Format: ONNX (CPU-optimized)
• Size: ~9 MB
• License: Apache 2.0
• Offline: Runs on $150 Chromebooks with no internet

---

⚡ Inference (Python + ONNX Runtime)

from onnxruntime import InferenceSession
import json

# Load
session = InferenceSession("mindspark-1.0.onnx")
with open("label_maps.json") as f:
    label_maps = json.load(f)

# Tokenize input (use BERT tokenizer with max_length=128, padding, truncation)
# Then run:
emotion_logits, mindset_logits, risk_logits = session.run(
    None,
    {"input_ids": input_ids, "attention_mask": attention_mask}
)