README.md

# AIMO 3 Local Submission

This folder contains `submission.ipynb`, a local-runtime version of the AIMO 3 competition notebook. Assets are read from this directory instead of Kaggle input paths.

## Folder layout

| Path | Description |
|------|-------------|
| `submission.ipynb` | Main inference notebook |
| `wheels.tar.gz` | Offline pip wheels archive (from Kaggle `aimo-3-utils`) |
| `GPT-OSS-120B/` | Actual model weights on disk (descriptive folder name) |
| `model/` | Path the notebook reads at runtime (generic name in code) |
| `test.csv` | Optional; used for local gateway testing |
| `setup/` | Created automatically when `wheels.tar.gz` is extracted |

## Model path

Store the checkpoint under **`GPT-OSS-120B/`** so the folder name identifies the weights (same layout as [Kaggle gpt-oss-120b](https://www.kaggle.com/models/danielhanchen/gpt-oss-120b/Transformers/default/1) or [Hugging Face gpt-oss-120b](https://huggingface.co/openai/gpt-oss-120b)).

The notebook uses a generic path in code:

```python
MODEL_PATH = BASE_DIR / "model"
```

So at runtime it looks for **`model/`**, not `GPT-OSS-120B/`. Point that path at your weights before running, for example:

- **Symlink / junction:** `model` → `GPT-OSS-120B`
- **Copy or rename:** put (or mirror) the files under `model/`

The descriptive name is for your layout; the notebook keeps the generic `model` folder name unchanged.

## Running locally

1. Set your Jupyter working directory to this folder.
2. Link or copy those files to `model/` so `MODEL_PATH` resolves correctly.
3. Optionally add `test.csv` for local evaluation.
4. Run all cells in `submission.ipynb`.

The paths cell prints `BASE_DIR`, `WHEELS_ARCHIVE`, `MODEL_PATH`, and `TEST_CSV` plus whether each path exists.

## Notes

- The notebook still targets a Linux GPU environment (CUDA, vLLM, `tar`, `kaggle_evaluation`), as on Kaggle.
- Do not use original or Metal builds when sourcing from Hugging Face; use the same Transformers layout as the Kaggle model bundle linked above.