{
  "fetched_at": "2026-10-22T10:21:51Z",
  "source_url": "https://www.kaggle.com/competitions/cs109b-hw3-contest-2023/data",
  "pages": [
    {
      "name": "data-description",
      "content": "For this problem, we will be working with a modified version of the [Kannada MNIST dataset](https://arxiv.org/pdf/1908.01242.pdf), which is a large database of handwritten digits in the indigenous language *Kannada*.\n\nThis dataset consists of 60,000 28x28 grayscale images of the ten digits, along with a test set of 10,000 images.\n\nFor this homework, we will simplify the problem by only using the digits labeled `0` and `1` owing to the similarity of the two symbols, and we will use a total of 1,200 samples for training (this includes the data you will use for validation).\n\nTo understand the dataset better, we recommend this [article](https://towardsdatascience.com/a-new-handwritten-digits-dataset-in-ml-town-kannada-mnist-69df0f2d1456) by Vinay Prabhu, the curator of the dataset.\n\nDATA FILES:\n\nHere's a brief description of the data files:\n- `kmnist_train.csv` is our training dataset and the last column contains our response class. The 784 other columns correspond to the pixel values of the 28x28 dimension image. Class 0 means a sample is the handwritten digit `0` and class `1` means a sample is the handwritten digit `1` in the Kannada written language. kmnist_train.csv has 1,200 samples.\n- `kmnist_test.csv` has a structure similar to `kmnist_train.csv`, however the class label column is NOT included in with the test set. `kmnist_test.csv` has 2,000 samples.\n",
      "isPublished": false,
      "hasIsPublished": false,
      "postTitle": "",
      "hasPostTitle": false,
      "mimeType": "",
      "hasMimeType": false
    },
    {
      "name": "Description",
      "content": "Kaggle leaderboard scores are accuracy scores calculated by Kaggle when you upload your predictions on the KMNIST TEST set.\n\n**SUBMISSION FORMAT REQUIREMENTS**:\n\n- `.csv` is the format that Kaggle will accept.\n\n- The uploaded `.csv` must contain 2 columns. The first column must be named `id` and needs to contain the test observation index numbers for each prediction, the second must be named `category` and needs to contain your class predictions (i.e. `0` or `1`) for each corresponding test observation index location. \n\n**IMPORTANT NOTES ABOUT SCORING**:\n\n- The **public leaderboard** on Kaggle displays your performance on only 30% of the test set.\n\n\n- After the competition is complete, the **private leaderboard** will show your performance on the remaining 70% of the test set.\n\n Points for this questions will be awarded based on the private leaderboard as follows.\n\n| Private Leaderboard Score | Points |\n|---------------------------|--------|\n| 0.945 ≤ score             | 5      |\n| 0.94 ≤ score < 0.945     | 4      |\n| 0.93  ≤ score < 0.94      | 3      |\n| 0.92  ≤ score < 0.93      | 2      |\n| 0.889 ≤ score < 0.92      | 1      |\n\n**BONUS:** An additional **5 bonus points** will be given to the **Top-3** competitors (as ranked on the hidden private leaderboard).\n",
      "isPublished": false,
      "hasIsPublished": false,
      "postTitle": "",
      "hasPostTitle": false,
      "mimeType": "",
      "hasMimeType": false
    },
    {
      "name": "Evaluation",
      "content": "The goal of this contest is to implement the regularization methods discussed in class in order to build an ANN model of a prescribed architecture that generalizes best on test data. The measuring metric is accuracy score. The winning model will be the one with the highest accuracy score on the hidden test set. NOTE: Highest accuracy on public leaderboard does not imply best model because the final winner is announced based on a hidden test set. This set will be made public only at the end of the competition.",
      "isPublished": false,
      "hasIsPublished": false,
      "postTitle": "",
      "hasPostTitle": false,
      "mimeType": "",
      "hasMimeType": false
    },
    {
      "name": "rules",
      "content": "Below are the rules governing this CS109B Kaggle competition. \n\nPlease also review the modelling restrictions **\"DOS\"** and **\"DON'TS\"** outlined in question 2.3.1 of the HW3 notebook for additional details on what IS and what IS NOT permitted.\n\n**RULES:**\n\n- Multiple Kaggle submissions are permitted (with a maximum of 20 submissions per-participant per-day), **just note that you will need to choose, on Kaggle, the ONE single submission to use for final scoring prior to the final HW3 submission deadline**, and **your submitted notebook MUST contain the matching code and model that generated your ONE chosen submission.**\n\n\n- **To repeat this point, the version of your final HW3 notebook submitted on Canvas MUST contain the same code and exact same model used to generate your ONE chosen Kaggle submission.** (TFs may rerun your notebook code to ensure comparable final leaderboard results.)\n\n\n- **Please do not manually label your submissions.** In other words, the labels should only be the outcome of your model.\n\n\n- **No external data are allowed, you MUST USE ONLY the KMNIST training and test data downloaded via the \"Data\" tab of the CS109B competition page.**\n\n\n- **Do not** create multiple accounts on Kaggle.\n",
      "isPublished": false,
      "hasIsPublished": false,
      "postTitle": "",
      "hasPostTitle": false,
      "mimeType": "",
      "hasMimeType": false
    }
  ]
}