{
  "fetched_at": "2026-10-19T08:18:17Z",
  "source_url": "https://www.kaggle.com/competitions/bench-tab-v2/data",
  "pages": [
    {
      "name": "data-description",
      "content": "The competition is constituted of 45 datasets used in the original paper. They fall under four categories: features may or may not contain categorical features and the target is either continuous or binary.\n\nThe main differences with datasets used in the paper are:\n- we only use one split (shuffle split with 60% for training and 20% for both public and private LB)\n- the binary target are encoded as 0 or 1\n\nFor each datasets we provide:\n- train data and target\n- test data\n- json files listing the features and an indicator describing if they are categorical or not.\n\nWe also provide more global dictionnaries describing each tasks. \n\nYou goal is to provide prediction for the test data of each task.",
      "isPublished": false,
      "hasIsPublished": false,
      "postTitle": "",
      "hasPostTitle": false,
      "mimeType": "",
      "hasMimeType": false
    },
    {
      "name": "Description",
      "content": "I downloaded the original datasets of the paper, performed a train / public / private split (60%, 20%, 20%) and set a simple lgbm as a Baseline for tree models for each of the task. Can you find ways to improve tree baselines or are you going to challenge their use altogether ?",
      "isPublished": false,
      "hasIsPublished": false,
      "postTitle": "",
      "hasPostTitle": false,
      "mimeType": "",
      "hasMimeType": false
    },
    {
      "name": "Evaluation",
      "content": "As a starter we chose to evaluate tasks with the most immediate metric using standard losses. Classification tasks are evaluated with log-loss and regression tasks with MSE. The metrics are then scaled at -1 with a naive baseline (standard deviation or average log_loss) and a catboost ensemble baseline as 0. Each task score is individually clipped at +/- 10 to avoid outliers tasks. The final metric should be interpretable as an average gain over a simple catboost ensemble baseline with respect to a naive baseline.  \n\n## Submission File\n\nFor each tasks and for each ID in the test set, you must predict a taget variable. The file should have the following format:\n\n    task, id, target\n    334_333464, 0, 1\n    334_355455, 1, 1\n    335_354359, 0, 8\n    etc.\n\nThe catboost ensemble baseline is provided to help you submit.",
      "isPublished": false,
      "hasIsPublished": false,
      "postTitle": "",
      "hasPostTitle": false,
      "mimeType": "",
      "hasMimeType": false
    },
    {
      "name": "rules",
      "content": "The dataset being public we wish that contestant try their best to improve the modelling instead of looking for the rest of the dataset. We can't realistically enforce any rule about cheating this way... If you intend to submit the public data ... just be aware that it generally won't look good anywhere. ",
      "isPublished": false,
      "hasIsPublished": false,
      "postTitle": "",
      "hasPostTitle": false,
      "mimeType": "",
      "hasMimeType": false
    },
    {
      "name": "abstract",
      "content": "We aim to replicate the paper \"Why do tree-based models still outperform deep learning on tabular data?\" as a competition. As the paper gather 59 datasets, it is in fact a 59 in 1 competition.  It is set as a benchmark to compare 'State of the art' tree models, Deep Learning models, or maybe newcomers such as AutoML solutions or KAN. \n\nYour goal is to provide a solution to outperform a simple boosted tree baseline. There is no limit on what you can use be it software or hardware.\n\nThe more general goal of the competition is to provide a benchmark so that newcomers can learn about best practices and more experienced people can evaluate advances in ML for tabular data. \n\n\n",
      "isPublished": false,
      "hasIsPublished": false,
      "postTitle": "",
      "hasPostTitle": false,
      "mimeType": "",
      "hasMimeType": false
    },
    {
      "name": "Acknowledgements",
      "content": "I'd like to thank [Léo Grinsztajn](https://www.kaggle.com/leogri), author of the initial paper, for discussion, and [Carl McBride Ellis](https://www.kaggle.com/carlmcbrideellis) for beta-testing this competition.",
      "isPublished": false,
      "hasIsPublished": false,
      "postTitle": "",
      "hasPostTitle": false,
      "mimeType": "",
      "hasMimeType": false
    }
  ]
}