{
  "fetched_at": "2026-10-20T06:39:28Z",
  "source_url": "https://www.kaggle.com/competitions/cartesius/data",
  "pages": [
    {
      "name": "data-description",
      "content": "In this competition, we make only the validation and the test set available.\n\n### Files\n\n*   **testset.json** - The test set\n*   **valset.json** - The validation set\n*   **sample_submission.csv** - A sample submission file with the correct format\n\n### Test & validation sets\n\nThese 2 files contain hand-crafted geometries, covering a wide variety of geometries, with different positions and scales.\n\nThere is one geometry per line, represented by a [WKT string](https://shapely.readthedocs.io/en/stable/manual.html#well-known-formats). You can load these geometries with `shapely` : \n\n`from shapely import wkt`  \n`polygon = wkt.loads(wkt_string)`\n\n### Which data to use for training\n\nThis is free : use whatever you want to train your model !\n\nThe easiest way is to use `spwk-cartesius` package, which provides a Pytorch dataset where polygons are randomly generated.",
      "isPublished": false,
      "hasIsPublished": false,
      "postTitle": "",
      "hasPostTitle": false,
      "mimeType": "",
      "hasMimeType": false
    },
    {
      "name": "Description",
      "content": "At [Spacewalk](https://www.spacewalk.tech/), our models deal with **geometries** a lot.\n\nBut it's **not trivial** for a NN to learn a **good representation** (good features) of the input geometries.\nAnd if the geometry is not well-represented, it **directly impacts the quality** of the results.\n\n---\n\nIn this competition, you will train a NN to learn good representation by training the model to predict various properties of the given polygon, such as: \n* **Area**\n* **Perimeter**\n* **Size** (on the x-axis and the y-axis)\n* **Concavity** (concavity is a **ratio** that represents *how much concave a polygon is*)\n* **Minimum clearance** (minimum clearance is **smallest distance** by which a point could be moved to produce an **invalid** geometry)\n* **Centroid**\n* **OMBR ratio** : (Oriented Minimum Bounding Rectangle ratio, representing *how close the polygon is to a rectangle*)\n* **OMBR Aspect ratio** (represents the wideness of the polygon)\n* **Opening ratio** (represents the ratio of deadspace (not wide enough) of the polygon)\n\n---\n\n📢 ***If you are reaching interesting results, [Spacewalk](https://www.spacewalk.tech/) is definitely interested in your profile for an ML engineer position! Please share your results through the [Discussion panel](https://www.kaggle.com/c/cartesius/discussion).***",
      "isPublished": false,
      "hasIsPublished": false,
      "postTitle": "",
      "hasPostTitle": false,
      "mimeType": "",
      "hasMimeType": false
    },
    {
      "name": "rules",
      "content": "### Submission Limits\n\nYou may submit a maximum of 20 entries per day.",
      "isPublished": false,
      "hasIsPublished": false,
      "postTitle": "",
      "hasPostTitle": false,
      "mimeType": "",
      "hasMimeType": false
    },
    {
      "name": "Get Started",
      "content": "To help you get started, please check out this [notebook](https://www.kaggle.com/nicolasswk/cartesius-transformer) !",
      "isPublished": false,
      "hasIsPublished": false,
      "postTitle": "",
      "hasPostTitle": false,
      "mimeType": "",
      "hasMimeType": false
    },
    {
      "name": "Evaluation",
      "content": "The test data provided in this competition contains a list of **hand-crafted geometries**. Use your trained model to predict the properties of each geometry !\n\n### Submission format\n\nThe submission file should be a CSV file with 2 fields :`id` and `value`.\n\n`id` has the following format : **`<geometry_id>_<task_name>`**\n\n*For tasks that uses several values (for example for centroid prediction, there is the x-value and the y-value), each value should appear on a different line, with an index in the task name.*\n\nExample of submission file : \n\n```\nid,value\n0_area,0.0\n0_perimeter,0.0\n0_size_0,0.0\n0_size_1,0.0\n0_convexity,0.0\n0_min_clear,0.0\n0_centroid_0,0.0\n0_centroid_1,0.0\n0_ombr_ratio,0.0\n0_aspect_ratio,0.0\n0_opening_ratio,0.0\n1_area,0.0\n1_perimeter,0.0\n...\n```\n\n### Scoring\n\nThis competition is evaluated on the weighted mean absolute error (WMAE):\n\n![](https://user-images.githubusercontent.com/78339260/145503169-ed2c3436-e89d-4da9-b5f1-e50665b5c7d3.jpg)\n\nWhere :\n* \\\\( n \\\\) is the number of rows\n* \\\\( \\hat{y}_i \\\\) is the predicted sales\n* \\\\( y_i \\\\) is the actual sales\n* \\\\( w_i \\\\) are weights",
      "isPublished": false,
      "hasIsPublished": false,
      "postTitle": "",
      "hasPostTitle": false,
      "mimeType": "",
      "hasMimeType": false
    }
  ]
}