{
  "id": 613580,
  "title": "Many teams have a score of 0.3 on the Leaderboard. How are these teams with the same score of 0.3 ranked?",
  "url": "/competitions/recodai-luc-scientific-image-forgery-detection/discussion/613580",
  "author_name": "",
  "post_date": "2025-10-28T04:39:50.327211800Z",
  "votes": 1,
  "comment_count": 5,
  "views": 0,
  "content": "<p>Is 0.3 the highest possible score?</p>",
  "messages": [
    {
      "id": "3307901",
      "postDate": "10/28/2025 04:39:50",
      "content": "<p>Is 0.3 the highest possible score?</p>",
      "rawMarkdown": "Is 0.3 the highest possible score?",
      "votes": null
    },
    {
      "id": "3307915",
      "postDate": "10/28/2025 05:44:55",
      "content": "<p>Because there is only one sample in sample_subsubmission. I'm also very puzzled. Theoretically, sample_submite.csv should have 1,100 ids</p>",
      "rawMarkdown": "Because there is only one sample in sample_subsubmission. I'm also very puzzled. Theoretically, sample_submite.csv should have 1,100 ids",
      "votes": null
    },
    {
      "id": "3308069",
      "postDate": "10/28/2025 14:03:14",
      "content": "<p>Hi, the maximum score is 1.0. Right now it's just hard to come up with a model that will give more than 0.3. The task is quite complicated and simple and even normal models will be worse</p>",
      "rawMarkdown": "Hi, the maximum score is 1.0. Right now it's just hard to come up with a model that will give more than 0.3. The task is quite complicated and simple and even normal models will be worse",
      "votes": null
    },
    {
      "id": "3308071",
      "postDate": "10/28/2025 14:03:46",
      "content": "<p>No, with a real inference, there are more images.</p>",
      "rawMarkdown": "No, with a real inference, there are more images.",
      "votes": null
    },
    {
      "id": "3308086",
      "postDate": "10/28/2025 14:34:44",
      "content": "<p>I think most models currently score around 0.3. Maybe the task is very complicated, and most models can’t beat 0.3. But I believe that soon, one model will surpass 0.3.</p>",
      "rawMarkdown": "I think most models currently score around 0.3. Maybe the task is very complicated, and most models can’t beat 0.3. But I believe that soon, one model will surpass 0.3.",
      "votes": null
    },
    {
      "id": "3360882",
      "postDate": "12/02/2025 17:20:50",
      "content": "<p>its because there is a significant difference in the type of data they gave us for training and the private dataset is much different, so the best performing models are the ones that just do nothing and predict ALL AUTHENTIC which gives then a score of 0.303 exactly, because that portion of the final testing private dataset is authentic, most can not go beyond that because the ones which dare to predict the non authentic samples get the authentic ones wrong and they lose 1 point for each false positive while only gaining a fraction of the point for each true positive they predict (because the F1 metric is cruel and only gives high score to precise masks that fit perfectly, which is very hard to pull off)</p>",
      "rawMarkdown": "its because there is a significant difference in the type of data they gave us for training and the private dataset is much different, so the best performing models are the ones that just do nothing and predict ALL AUTHENTIC which gives then a score of 0.303 exactly, because that portion of the final testing private dataset is authentic, most can not go beyond that because the ones which dare to predict the non authentic samples get the authentic ones wrong and they lose 1 point for each false positive while only gaining a fraction of the point for each true positive they predict (because the F1 metric is cruel and only gives high score to precise masks that fit perfectly, which is very hard to pull off)",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3307915,
      "author_name": "bent1e",
      "author_url": "",
      "post_date": "10/28/2025 05:44:55",
      "content": "<p>Because there is only one sample in sample_subsubmission. I'm also very puzzled. Theoretically, sample_submite.csv should have 1,100 ids</p>",
      "votes": null,
      "replies": [
        {
          "id": 3308071,
          "author_name": "antonoof",
          "author_url": "",
          "post_date": "10/28/2025 14:03:46",
          "content": "<p>No, with a real inference, there are more images.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 3308069,
      "author_name": "antonoof",
      "author_url": "",
      "post_date": "10/28/2025 14:03:14",
      "content": "<p>Hi, the maximum score is 1.0. Right now it's just hard to come up with a model that will give more than 0.3. The task is quite complicated and simple and even normal models will be worse</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 3308086,
      "author_name": "nobita7",
      "author_url": "",
      "post_date": "10/28/2025 14:34:44",
      "content": "<p>I think most models currently score around 0.3. Maybe the task is very complicated, and most models can’t beat 0.3. But I believe that soon, one model will surpass 0.3.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 3360882,
      "author_name": "mrearthworm",
      "author_url": "",
      "post_date": "12/02/2025 17:20:50",
      "content": "<p>its because there is a significant difference in the type of data they gave us for training and the private dataset is much different, so the best performing models are the ones that just do nothing and predict ALL AUTHENTIC which gives then a score of 0.303 exactly, because that portion of the final testing private dataset is authentic, most can not go beyond that because the ones which dare to predict the non authentic samples get the authentic ones wrong and they lose 1 point for each false positive while only gaining a fraction of the point for each true positive they predict (because the F1 metric is cruel and only gives high score to precise masks that fit perfectly, which is very hard to pull off)</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3307901": "Is 0.3 the highest possible score?",
    "3307915": "Because there is only one sample in sample_subsubmission. I'm also very puzzled. Theoretically, sample_submite.csv should have 1,100 ids",
    "3308069": "Hi, the maximum score is 1.0. Right now it's just hard to come up with a model that will give more than 0.3. The task is quite complicated and simple and even normal models will be worse",
    "3308071": "No, with a real inference, there are more images.",
    "3308086": "I think most models currently score around 0.3. Maybe the task is very complicated, and most models can’t beat 0.3. But I believe that soon, one model will surpass 0.3.",
    "3360882": "its because there is a significant difference in the type of data they gave us for training and the private dataset is much different, so the best performing models are the ones that just do nothing and predict ALL AUTHENTIC which gives then a score of 0.303 exactly, because that portion of the final testing private dataset is authentic, most can not go beyond that because the ones which dare to predict the non authentic samples get the authentic ones wrong and they lose 1 point for each false positive while only gaining a fraction of the point for each true positive they predict (because the F1 metric is cruel and only gives high score to precise masks that fit perfectly, which is very hard to pull off)"
  },
  "source": "meta"
}