A safe, guardrailed AI tutor specifically designed for teaching statistics and probability to high school students (ages 14–18). Built for the Oklahoma Department of Education project. Runs offline on phones via PocketPal.

🎯 What This Model Does

StatBot is a fine-tuned vision-language model that:

📊 Teaches statistics — mean, median, distributions, hypothesis testing, probability, regression, and more
🖼️ Reads images — graphs, charts, histograms, textbook pages, handwritten formulas
🛡️ Refuses unsafe content — off-topic questions, jailbreaks, inappropriate requests
💙 Supports students in distress — empathy + counselor referral
🎓 Uses hint-first teaching — guides students rather than just giving answers

📁 Files in This Repo

File	Size	What It Is
`Qwen3VL-2B-StatBot-Q8_0.gguf`	1.71 GB	Main model (Q8_0 quantized)
`Qwen3VL-2B-StatBot-mmproj.gguf`	782 MB	Vision projector (needed for images)

⚠️ Both files required for full image + text capability!

📱 How to Use on Your Phone (PocketPal)

Install PocketPal from your app store
Open PocketPal → Models → Search HuggingFace
Search: Anudeeprao1/Qwen3VL-2B-StatBot_Instruct
Download BOTH files:
- ✅ Qwen3VL-2B-StatBot-Q8_0.gguf
- ✅ Qwen3VL-2B-StatBot-mmproj.gguf
Tap Load Model
Start asking statistics questions or upload a graph!

💻 How to Use with Ollama

ollama run hf.co/Anudeeprao1/Qwen3VL-2B-StatBot_Instruct

💻 How to Use with llama.cpp

llama-server -hf Anudeeprao1/Qwen3VL-2B-StatBot_Instruct

Then open http://localhost:8080 in browser.

💡 Example Questions

Text Only

"What is the difference between mean and median?"
"Explain the central limit theorem to me"
"How do I calculate standard deviation step by step?"
"What is a p-value?"

With Images

Upload a histogram → "What type of distribution is this?"
Upload a textbook page → "Explain this formula"
Upload a scatter plot → "What relationship does this show?"
Upload a bar chart → "Describe this data"

🛡️ Safety Features

StatBot is designed with 7 categories of guardrails:

Category	Behavior
Off-topic questions	Refuses politely, redirects to stats
Inappropriate content	Refuses firmly
Identity probing	Deflects, never reveals internals
Jailbreak attempts	Resists prompt injection
Student distress	Shows empathy, refers to counselor
Privacy questions	Confirms FERPA + COPPA design
Educational audits	Provides compliance info

⚙️ Training Details

Parameter	Value
Base model	unsloth/Qwen3-VL-2B-Instruct
Fine-tuning method	LoRA SFT (Unsloth)
LoRA rank	16
Quantization	Q8_0 GGUF
Training samples	7,202
Guardrail samples	720 (7 categories × 3 repetitions)
Epochs	3
Learning rate	2e-4 (cosine schedule)
Effective batch size	16
Max sequence length	2048
Hardware	NVIDIA RTX 4500 Ada (24 GB)
Framework	Unsloth + TRL + Transformers

📊 Performance

Final training loss: ~0.09
Final validation loss: ~0.09
Perplexity: ~1.1
No overfitting detected (train ≈ val loss)
Vision capability: Full multimodal support via mmproj

🎓 Intended Audience

✅ High school statistics students (ages 14–18)
✅ Oklahoma high school teachers (classroom aid)
✅ Statistics tutors needing safe AI assistance
✅ Educational researchers studying AI guardrails
❌ NOT for general-purpose chat (refuses non-stats)
❌ NOT for elementary/middle school content

🏛️ Compliance & Safety

FERPA compliant design (no student data storage)
COPPA compliant for under-13 protections
Oklahoma DoE project deliverable
No cross-session memory (every chat is private)
Crisis-aware (refers to 988 Lifeline + adults)

📚 Try a Sample Conversation

User: "What is standard deviation?"

StatBot: "Great question! 🎯 Standard deviation tells us how spread out data is from the average...

Think of test scores: if everyone scored close to the average, SD is small. If scores are all over the place, SD is large.

Formula: σ = √(Σ(x - μ)² / N)

Want me to walk through an example?"

User: "What is the capital of France?"

StatBot: "I'm StatBot, your statistics tutor designed for educational purposes only! I can only help with high school statistics topics. Try asking me about mean, median, probability, or graphs! Have a good day. 📊"

⚠️ Limitations

Only English language
Only statistics/probability topics
Best for high school level (ages 14–18)
Requires both GGUF files for image input
Q8_0 quantization (high quality but ~2 GB)

📄 License

Apache 2.0 — Free for educational and research use.

👤 Author

Anudeep — Oklahoma Department of Education AI Tutoring Project

For questions or collaboration: open a discussion on this repo!

🏷️ Citation

If you use this model in research, please cite:

bibtex

@misc{statbot2026,
  author       = {Anudeep},
  title        = {StatBot: A Guardrailed Vision-Language Tutor for High School Statistics},
  year         = {2026},
  publisher    = {HuggingFace},
  url          = {https://huggingface.co/Anudeeprao1/Qwen3VL-2B-StatBot_Instruct}
}

🙏 Acknowledgments

Unsloth — for the 2x faster fine-tuning framework
Qwen team — for the excellent base VL model
Oklahoma DoE — for funding this educational AI research
PocketPal — for offline mobile inference support