{
  "id": 614650,
  "title": "Clarification on test data distribution and final evaluation set",
  "url": "/competitions/recodai-luc-scientific-image-forgery-detection/discussion/614650",
  "author_name": "",
  "post_date": "2025-11-05T12:16:59.474000",
  "votes": 1,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Hi, I have a question about the test data used on the current public leaderboard.\nCould you please confirm whether the distribution of the current test set is representative of the final private test set used for evaluation?</p>\n<p>Specifically:</p>\n<p>Will the final test set only include the same categories of images as the current test set (for example, only corn-related images), or might it contain additional/unseen types (e.g., potato images or others)?</p>\n<p>My approach relies on K-means clustering based on image features, so I want to make sure the distribution consistency holds for the final evaluation.</p>\n<p>Thank you for clarifying!</p>",
  "messages": [
    {
      "id": 3311616,
      "postDate": "2025-11-05T12:16:59.473Z",
      "content": "<p>Hi, I have a question about the test data used on the current public leaderboard.\nCould you please confirm whether the distribution of the current test set is representative of the final private test set used for evaluation?</p>\n<p>Specifically:</p>\n<p>Will the final test set only include the same categories of images as the current test set (for example, only corn-related images), or might it contain additional/unseen types (e.g., potato images or others)?</p>\n<p>My approach relies on K-means clustering based on image features, so I want to make sure the distribution consistency holds for the final evaluation.</p>\n<p>Thank you for clarifying!</p>",
      "rawMarkdown": "Hi, I have a question about the test data used on the current public leaderboard.\nCould you please confirm whether the distribution of the current test set is representative of the final private test set used for evaluation?\n\nSpecifically:\n\nWill the final test set only include the same categories of images as the current test set (for example, only corn-related images), or might it contain additional/unseen types (e.g., potato images or others)?\n\nMy approach relies on K-means clustering based on image features, so I want to make sure the distribution consistency holds for the final evaluation.\n\nThank you for clarifying!",
      "votes": 1
    },
    {
      "id": 3311625,
      "postDate": "2025-11-05T12:35:11.610Z",
      "content": "<p>I think corn only accounts for a very tiny portion. This competition requires us to train a model that can adapt to the vast majority of images. So I believe the final test set will contain some unseen images, just like 45.png in the current test set (which is a chart).Good luck!</p>",
      "rawMarkdown": "I think corn only accounts for a very tiny portion. This competition requires us to train a model that can adapt to the vast majority of images. So I believe the final test set will contain some unseen images, just like 45.png in the current test set (which is a chart).Good luck!",
      "replies": [
        {
          "id": 3311634,
          "postDate": "2025-11-05T12:49:59.950Z",
          "content": "<p>I understand that 45.png in the current test set represents a completely new image type, and I also know that the images used for the public leaderboard are not accessible to participants.</p>\n<p>What I would like to clarify is:</p>\n<p>Is the current public leaderboard test set sampled uniformly or representatively from the final private test set?</p>\n<p>In other words, are their image distributions the same?</p>\n<p>Can the public leaderboard score be considered a meaningful indicator of final performance on the private test set?</p>\n<p>I’m asking this because my method is based on K-means clustering, and I’d like to confirm whether its behavior and assumptions will remain valid on the final evaluation set.😀</p>",
          "rawMarkdown": "I understand that 45.png in the current test set represents a completely new image type, and I also know that the images used for the public leaderboard are not accessible to participants.\n\nWhat I would like to clarify is:\n\nIs the current public leaderboard test set sampled uniformly or representatively from the final private test set?\n\nIn other words, are their image distributions the same?\n\nCan the public leaderboard score be considered a meaningful indicator of final performance on the private test set?\n\nI’m asking this because my method is based on K-means clustering, and I’d like to confirm whether its behavior and assumptions will remain valid on the final evaluation set.😀"
        }
      ]
    },
    {
      "id": 3311624,
      "postDate": "2025-11-05T12:34:26.043Z",
      "rawMarkdown": "",
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 3311625,
      "author_name": "耶✌",
      "author_url": "",
      "post_date": "2025-11-05T12:35:11.610000",
      "content": "<p>I think corn only accounts for a very tiny portion. This competition requires us to train a model that can adapt to the vast majority of images. So I believe the final test set will contain some unseen images, just like 45.png in the current test set (which is a chart).Good luck!</p>",
      "votes": 0,
      "replies": [
        {
          "id": 3311634,
          "author_name": "",
          "author_url": "",
          "post_date": "2025-11-05T12:49:59.950000",
          "content": "<p>I understand that 45.png in the current test set represents a completely new image type, and I also know that the images used for the public leaderboard are not accessible to participants.</p>\n<p>What I would like to clarify is:</p>\n<p>Is the current public leaderboard test set sampled uniformly or representatively from the final private test set?</p>\n<p>In other words, are their image distributions the same?</p>\n<p>Can the public leaderboard score be considered a meaningful indicator of final performance on the private test set?</p>\n<p>I’m asking this because my method is based on K-means clustering, and I’d like to confirm whether its behavior and assumptions will remain valid on the final evaluation set.😀</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 3311624,
      "author_name": "",
      "author_url": "",
      "post_date": "2025-11-05T12:34:26.043000",
      "content": "",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3311616": "Hi, I have a question about the test data used on the current public leaderboard.\nCould you please confirm whether the distribution of the current test set is representative of the final private test set used for evaluation?\n\nSpecifically:\n\nWill the final test set only include the same categories of images as the current test set (for example, only corn-related images), or might it contain additional/unseen types (e.g., potato images or others)?\n\nMy approach relies on K-means clustering based on image features, so I want to make sure the distribution consistency holds for the final evaluation.\n\nThank you for clarifying!",
    "3311625": "I think corn only accounts for a very tiny portion. This competition requires us to train a model that can adapt to the vast majority of images. So I believe the final test set will contain some unseen images, just like 45.png in the current test set (which is a chart).Good luck!",
    "3311624": ""
  }
}