First Python

• 

  Seatmate classroom. Start wherever you are. One idea, one example, one checkpoint — then it waits.

Inspiration

A lot of teens cannot afford a private tutor. Chatbots dump the answer, which feels helpful and teaches almost nothing. I wanted the empty seat next to you: a classroom that waits. One idea, one tiny example, one checkpoint.

What it does

Seatmate is a Socratic tutor for coding, math, and science. Ask for a lesson, a multi-week course, or a practice test. If you miss, it hints first. It will not complete a graded school test for you. Math and science facts get a live Wolfram|Alpha check so it does not invent numbers.

Live demo: https://llm-teacher-yvxl.onrender.com
(The free host sleeps when idle — the first load can take about a minute.)

How we built it

I wrote a 1,712-example Socratic dataset (plus 152 validation rows), then QLoRA-tuned Qwen3-14B for 1 epoch (107 steps, loss 2.92 → 0.39). I merged the adapter to full weights, put it on Hugging Face, and serve it through Featherless. A small FastAPI classroom streams replies and attaches Wolfram only at inference time — not as training data.

Code: https://github.com/zivzancoeli-commits/LLM-Teacher
Model: https://huggingface.co/EZiv/llm-teacher-qwen3-14b

Challenges we ran into

GPU merge and upload are large. Featherless Chat cannot host a private model, so the weights had to be public. Qwen3 “thinking” tokens first showed up as empty classroom replies until the app read that stream. The demo sleeps on the free host, so judges should wait on the first visit.

Accomplishments that we're proud of

The tutor is a real fine-tune, not a stock model with a prompt glued on. It teaches, then waits. It refuses to take a graded test. I shipped a public classroom a teenager can open in a browser.

What we learned

A small, careful dataset beats repeating the same rows. Serving a 14B fine-tune is a different problem from training it. Access matters more than sounding smart.

What's next for Seatmate

Keep the Socratic style, tighten math/science checks, and make the first-load wait shorter so more students can actually use it.

Built With

• fastapi
• featherless
• hugging-face
• machine-learning
• python
• qlora
• qwen
• render
• unsloth