AIMO3_Prompt_Avoiding-Frequent-Errors_GoogleColab

AIMO3 Prompt Avoiding Frequent Errors (Colab Version)

This repository contains the Google Colab version of the Kaggle AIMO3 competition notebook. The original Kaggle submission code was modified to run smoothly on Google Colab (A100).

---

1. Overview

This notebook processes math problems from a CSV file (or a dummy data table) one by one and outputs the answers.

• Core Logic: The main calculation logic is exactly the same as the original Kaggle submission code.
• Input: Reads a question list from a CSV file or an internal table.
• Output: Saves and prints answers sequentially with real-time accuracy checks.

---

2. Key Differences from Kaggle Notebook

We removed Kaggle-dependent codes and optimized the notebook for the Google Colab environment.

🔴 Removed (Kaggle Specific)

• kaggle_evaluation library and API server code.
• Kaggle local paths (e.g., /kaggle/input/..., /kaggle/usr/lib/...).

🟢 Added / Changed (Colab Specific)

• Model Path: Changed from local Kaggle path to Hugging Face Model ID (openai/gpt-oss-120b).
• Data Input: Replaced the Kaggle submission API with a simple pandas loop to process CSV files or dummy data.
• Environment: Added library installations for Colab (VLLM, Transformers, Torch, etc.) and HF_TOKEN setup using Colab Secrets.

---

3. How to Run on Colab

1. Open your .ipynb file in Google Colab.
2. Select H100 or A100 (High-RAM) GPU.
3. Set your Hugging Face Token in Colab Secrets as HF_TOKEN, or enter your token directly in the notebook.
4. Run the notebook cells from the top.

GPU and Memory Settings

• H100 / A100 Settings: In the original Kaggle notebook, kv_cache_dtype = 'fp8_e4m3' was used. If you can select an H100 GPU on Colab, this setting is highly recommended. However, because Google Colab did not allocate an H100 to me during testing, this public code is currently configured with kv_cache_dtype = 'auto' to run on A100 (High-RAM).