MAP - Charting Student Math Misunderstandings

Competition

Solution Writeup

Requirements

Hardware

Memory: 144GB or more

GPU: 96GB or more

NVIDIA RTX PRO 6000 Blackwell Workstation Edition was used.

Software

Check docker/Dockerfile

docker build -t kaggle-map -f docker/Dockerfile docker/
docker run -it --rm --gpus all --ipc=host \
  -v $(pwd)/..:/workspace \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  kaggle-map

Directory structure should be like this:

.
├── data
│   ├── map-charting-student-math-misunderstandings.zip
│   ├── sample_submission.csv
│   ├── test.csv
│   └── train.csv
├── artifacts
│   ├── dtrainval.csv
│   └── dtrainval_qwen3_235b_a22b_thinking_2507_fp8.parquet
└── kaggle-map-charting-student-math-misunderstandings
    ├── LICENSE
    ├── README.md
    ├── configs
    └── ...

Prepare Data

Make 5-fold splits of the data.

python scripts/prepare_data.py

Generate responses using Qwen3-235B-A22B-Thinking-2507-FP8. This requires A100-SXM4-80GB x4 and takes about 1h45m. Use the result in data/dtrainval_qwen3_235b_a22b_thinking_2507_fp8.parquet.

python scripts/prepare_response.py

Training

Check train.sh for details.

Inference

Check convert.sh for details.

Kaggle Notebook