Kaggle2026HackathonsMLops
AI Mathematical Olympiad: Progress Prize 3
Getting an LLM to solve olympiad-level maths inside an offline Kaggle kernel. A QLoRA fine-tune of Qwen3-30B-A3B, vLLM inference with tool-integrated reasoning, and majority voting.
Qwen3-30B-A3BUnslothQLoRAvLLM
Status: Draft.
Objective
Solve olympiad-level maths problems with an LLM in an offline Kaggle kernel.
Approach
- Fine-tune: QLoRA on Qwen3-30B-A3B with Unsloth.
- Inference: vLLM with tool-integrated reasoning. The model writes code to check its own working.
- Majority voting over several sampled solutions.
ℹ️ Info: the kernel has no internet access, so a separate dependency-bundling notebook packages every wheel and model file ahead of time. Getting the environment to run offline is half the competition.
Where it's at
- ✅ Dependency-bundling notebook for the no-internet kernel
- 🔄 QLoRA fine-tune
- 🔄 vLLM inference with tool use and voting
Posts
No posts for this project yet. Come back later.