Eedi 5th solution

This repository contains the training code for the 5th place solution in the kaggle Eedi competition. Solution page

Setup

Download competition data

$ kaggle competitions download -c eedi-mining-misconceptions-in-mathematics
$ unzip eedi-mining-misconceptions-in-mathematics.zip -d input

Build python environment

• Install uv → https://docs.astral.sh/uv/#getting-started
• Install and activate environment

   $ uv sync
   $ . .venv/bin/activate
  

Set OpenRouter API key (this is only required for generation of additional data)

• set your openrouter api key in .env

   $ cp .env_sample .env
  

login wandb

$ wandb login

Run

• Run all pipeline
  With default parameters(This is almost identical to what we used in our best sub), biencoder training can be done on a 1x4090 GPU. Training the listwise reranker requires an 1xA100 (VRAM: 80GB)

   $ bash run_all.sh
  

• You can also execute each step by downloading files that contain the intermediate progress saved for each step.

PS: We don't intend to maintain this repository.