Matus is a fine-tuned Llama 3.2 3B model built for K-12 and early college math tutoring. It is the backbone of Project Matus — an open-source, locally-run AI tutoring system designed for students whose ways of knowing have historically been left out of math education.

Core design principles:

Never gives answers directly — scaffolds student thinking through questions
Recognizes valid alternative mathematical frameworks before correcting
Protects productive struggle — does not recalibrate downward because a student is frustrated
Culturally responsive — code-switching, family methods, and non-Western approaches are assets
Neurodivergence-aware — short responses and flat affect are not treated as disengagement
No data collection. No cloud. No subscriptions. Runs entirely on local hardware.

What It Runs On

Format: GGUF Q4_K_M quantization
Compatible with: llama.cpp, Ollama, LM Studio, any GGUF-compatible runtime
Minimum hardware: 8 GB RAM, any CPU (Intel, AMD, Apple Silicon)
Recommended: 16 GB RAM for comfortable performance
Latency: 30–90 seconds per response on CPU-only hardware. Faster on Apple Silicon or GPU.

How To Run It

One command with Ollama:

ollama run hf.co/TushaeBXN/matus-3b:Q4_K_M

One command with llama.cpp:

llama-server -hf TushaeBXN/matus-3b:Q4_K_M

With the full Project Matus system (auto-downloads model):

bash

git clone https://github.com/TushaeBXN/project-matus.git
cd project-matus
pip install -r requirements.txt
./start.sh

K-12 Math Tutor (with simulated student profiles):

bash

./boot_server.sh
python3 tutor/main.py --student james --role teacher

Training

Base model: Llama 3.2 3B Instruct (unsloth/Llama-3.2-3B-Instruct)
Method: LoRA fine-tuning (r=16, lora_alpha=32)
Dataset: 250 curated examples — identity data, conversational responses, K-12 and early college math tutoring scenarios across 9 domains
Training hardware: NVIDIA RTX A6000 (48GB VRAM) via RunPod
Framework: Unsloth + TRL SFTTrainer
Epochs: 3

Evaluation Results

Evaluated on 15 held-out math tutoring problems against a prompt-only baseline:

Metric	Baseline	Matus 3B	Target
Answer giveaway rate	0.0%	0.0%	<5%
Scaffolding quality	70.0%	72.7%	>70%
Conceptual accuracy	23.3%	32.6%	>70%
Ends with question	67%	86%	—

Behavior improvements over baseline: honor_struggle, affirm_partial, recognize_fallacy, counterexample, explain_composition all moved from 0% to 100%.

Full evaluation report: docs/preliminary_data.md

What It Won't Do

No data collection — nothing leaves your machine
No cloud dependency — runs fully offline after download
No answer giveaways — designed to scaffold, not solve
No biometric input — affect detection is text-based only
No diagnostic labeling — student memory stores behavioral observations, never deficit labels

Project

Part of Project Matus — an open-source K-12 math tutoring platform with:

Thought-token reasoning pipeline (internal reasoning hidden from student)
Affect detection (Tier 1/2/3 escalation) grounded in transformative SEL
Cross-session student memory (detects students who disengage across multiple sessions)
Epistemically just session logging (annotation-ready for researcher review)
5 simulated student profiles for testing and co-design

GitHub: github.com/TushaeBXN/project-matus

License

Apache 2.0 — free to use, modify, and distribute with attribution. Base model license: Llama 3.2 Community License (Meta).