ACLASCollege/neuro-edu-core-v3

ACLAS Neuro-Edu Core v3.0 — Tiny-Cognition (ATC) v2

The on-device neural kernel powering every Neuro-Edu student agent: a tiny multi-layer perceptron that predicts cognitive absorption probability from a student state vector.

Model

• Architecture: Input(5) → Dense(16) → ReLU → Dense(8) → ReLU → Dense(1) → Sigmoid — 241 parameters, pure NumPy, no dependencies
• Input features [complexity, attention, skill_match, fatigue, prior_knowledge], each in [0, 1]
• Output: absorption_probability in (0, 1)
• Training: 4,000 synthetic samples, 30 epochs SGD (lr 0.005, MSE loss, gradient clipping), seed 20261018 — holdout MAE 0.012

Files

• atc-v3.npz — weights (w0/b0 … w2/b2)
• config.json — architecture and training config
• load_model.py — standalone NumPy loader (zero dependencies)

Usage

from load_model import TinyCognition
model = TinyCognition("atc-v3.npz")
# [complexity, attention, skill_match, fatigue, prior_knowledge]
print(model.predict([0.6, 0.8, 0.7, 0.2, 0.5]))  # -> ~0.92

Resources

• GitHub: aclascollege/neuro-edu
• Live Demo: Interactive Neuro-Edu Dashboard

Citation

@techreport{aclas_neuro_edu_2026,
  title={Neuro-Edu: A Cognitive Science-Driven Framework for Large Language Model Alignment in Educational Sandbox Simulation},
  author={Atlanta College of Liberal Arts and Sciences},
  year={2026},
  institution={ACLAS College},
  doi={10.5281/zenodo.19743807},
  url={https://doi.org/10.5281/zenodo.19743807}
}