This repository contains yemaoye_submission.ipynb, a local-runtime version of my Kaggle AIMO 3 competition notebook.
The core inference logic is kept consistent with my final Kaggle submission. Kaggle-specific input paths have been replaced with local paths relative to this repository.
Folder layout
Path
Description
yemaoye_submission.ipynb
Main inference notebook
wheels.tar.gz
Offline pip wheels archive used for environment setup
model/
Path the notebook reads at runtime for GPT-OSS-120B weights
test.csv
Optional; used for local gateway testing
setup/
Created automatically when wheels.tar.gz is extracted
NOTICE.md
Attribution and acknowledgement information
LICENSE
License file
Model path
The notebook uses the following generic model path:
MODEL_PATH = BASE_DIR / "model"
This repository does not duplicate the full GPT-OSS-120B weights because of their large size.
Before running, point model/ to the GPT-OSS-120B weights, for example:
Symlink / junction: model → local GPT-OSS-120B directory
Copy or rename: put the model files under model/
Or modify MODEL_PATH inside the notebook
Running locally
Set your Jupyter working directory to this repository.
Make sure wheels.tar.gz is in the repository root.
Link or copy GPT-OSS-120B weights to model/.
Optionally add test.csv for local gateway testing.
Run all cells in yemaoye_submission.ipynb.
The notebook is intended for a Linux GPU environment similar to Kaggle, with CUDA, vLLM, tar, and the required offline wheels.
Notes
The full GPT-OSS-120B model weights are not included here.
setup/ and setup/tiktoken_encodings/ are generated during environment setup.
Acknowledgements
My solution is largely built upon the excellent public baseline notebook shared by Andreas Bisiadis. I sincerely thank Andreas Bisiadis and other open-source contributors for their generous sharing and strong engineering work.
I also thank Kaggle and the competition host team for organizing the AI Mathematical Olympiad - Progress Prize 3 competition.
