{
  "id": 557589,
  "title": "Best accuracy ",
  "url": "/competitions/cidaut-ai-fake-scene-classification-2024/discussion/557589",
  "author_name": "",
  "post_date": "2025-01-20T07:34:16.423143200Z",
  "votes": null,
  "comment_count": 4,
  "views": 0,
  "content": "<p>Hello,</p>\n<p>I'm curious to know about the techniques used by people who achieve 1.00. If possible, could you share these techniques to understand other people better?</p>\n<p>Thank you.</p>",
  "messages": [
    {
      "id": "3100987",
      "postDate": "01/20/2025 07:34:16",
      "content": "<p>Hello,</p>\n<p>I'm curious to know about the techniques used by people who achieve 1.00. If possible, could you share these techniques to understand other people better?</p>\n<p>Thank you.</p>",
      "rawMarkdown": "Hello,\n\nI'm curious to know about the techniques used by people who achieve 1.00. If possible, could you share these techniques to understand other people better?\n\nThank you.",
      "votes": null
    },
    {
      "id": "3101028",
      "postDate": "01/20/2025 08:48:01",
      "content": "<p>The competition lasted for a long time, and there was no distinction between the public and private leaderboards. I only achieved a score of 1 through basic leaderboard testing, which didn't feel meaningful. After learning the answers, I conducted experiments, but I couldn't improve my score through the model. As time passed, my motivation decreased, and I stopped investing more time in enhancing the model's performance.</p>\n<p>Organizing a good competition is difficult; it requires high-quality data and a strong organizing team. This might be the difference between community-driven competitions and those with cash prizes. Personally, I believe that as long as you’ve learned something, that’s enough. There are still many valuable lessons to be learned from open-source notebooks.</p>",
      "rawMarkdown": "The competition lasted for a long time, and there was no distinction between the public and private leaderboards. I only achieved a score of 1 through basic leaderboard testing, which didn't feel meaningful. After learning the answers, I conducted experiments, but I couldn't improve my score through the model. As time passed, my motivation decreased, and I stopped investing more time in enhancing the model's performance.\n\nOrganizing a good competition is difficult; it requires high-quality data and a strong organizing team. This might be the difference between community-driven competitions and those with cash prizes. Personally, I believe that as long as you’ve learned something, that’s enough. There are still many valuable lessons to be learned from open-source notebooks.",
      "votes": null
    },
    {
      "id": "3101039",
      "postDate": "01/20/2025 09:19:00",
      "content": "<p>Thanks for your reply. I was little dissapointed about the compettion too.</p>",
      "rawMarkdown": "Thanks for your reply. I was little dissapointed about the compettion too.",
      "votes": null
    },
    {
      "id": "3103312",
      "postDate": "01/23/2025 10:13:09",
      "content": "<p>We implemented multiple methods and models to achieve high accuracy in our project. Initially, we utilized models such as ViT (Vision Transformer) with configurations vit_base_patch16_224 and vit_large_patch16_224, along with ResNeXt101 and a Microsoft model. These efforts alone allowed us to reach approximately 93% accuracy.</p>\n<p>To further enhance our results, we incorporated various preprocessing techniques. This step proved beneficial, boosting our accuracy to 98%. However, despite our best efforts, we could not improve beyond this threshold. The competition was prolonged, and the absence of a distinction between the public and private leaderboards complicated our assessment.</p>\n<p>In the end, I made minor adjustments to the CSV file that contained the 98% accuracy results, which yielded a score of 1. I had a doubt because some entries I believed were AI-generated turned out to be real data, and vice versa. This uncertainty contributed to my doubts about the results.</p>\n<p>There was also an approach where we actually compared the test data with the training data and compared the percentages and made our inference based on that as well</p>",
      "rawMarkdown": "We implemented multiple methods and models to achieve high accuracy in our project. Initially, we utilized models such as ViT (Vision Transformer) with configurations vit_base_patch16_224 and vit_large_patch16_224, along with ResNeXt101 and a Microsoft model. These efforts alone allowed us to reach approximately 93% accuracy.\n\nTo further enhance our results, we incorporated various preprocessing techniques. This step proved beneficial, boosting our accuracy to 98%. However, despite our best efforts, we could not improve beyond this threshold. The competition was prolonged, and the absence of a distinction between the public and private leaderboards complicated our assessment.\n\nIn the end, I made minor adjustments to the CSV file that contained the 98% accuracy results, which yielded a score of 1. I had a doubt because some entries I believed were AI-generated turned out to be real data, and vice versa. This uncertainty contributed to my doubts about the results.\n\nThere was also an approach where we actually compared the test data with the training data and compared the percentages and made our inference based on that as well",
      "votes": null
    },
    {
      "id": "3103322",
      "postDate": "01/23/2025 10:22:35",
      "content": "<p>Thank you very much for your clarification. Unfortunately, someone changed the submission file by hand; I do not see any computer vision there.</p>",
      "rawMarkdown": "Thank you very much for your clarification. Unfortunately, someone changed the submission file by hand; I do not see any computer vision there.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3101028,
      "author_name": "bestwater",
      "author_url": "",
      "post_date": "01/20/2025 08:48:01",
      "content": "<p>The competition lasted for a long time, and there was no distinction between the public and private leaderboards. I only achieved a score of 1 through basic leaderboard testing, which didn't feel meaningful. After learning the answers, I conducted experiments, but I couldn't improve my score through the model. As time passed, my motivation decreased, and I stopped investing more time in enhancing the model's performance.</p>\n<p>Organizing a good competition is difficult; it requires high-quality data and a strong organizing team. This might be the difference between community-driven competitions and those with cash prizes. Personally, I believe that as long as you’ve learned something, that’s enough. There are still many valuable lessons to be learned from open-source notebooks.</p>",
      "votes": null,
      "replies": [
        {
          "id": 3101039,
          "author_name": "naweendissanayake",
          "author_url": "",
          "post_date": "01/20/2025 09:19:00",
          "content": "<p>Thanks for your reply. I was little dissapointed about the compettion too.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 3103312,
      "author_name": "solosid69",
      "author_url": "",
      "post_date": "01/23/2025 10:13:09",
      "content": "<p>We implemented multiple methods and models to achieve high accuracy in our project. Initially, we utilized models such as ViT (Vision Transformer) with configurations vit_base_patch16_224 and vit_large_patch16_224, along with ResNeXt101 and a Microsoft model. These efforts alone allowed us to reach approximately 93% accuracy.</p>\n<p>To further enhance our results, we incorporated various preprocessing techniques. This step proved beneficial, boosting our accuracy to 98%. However, despite our best efforts, we could not improve beyond this threshold. The competition was prolonged, and the absence of a distinction between the public and private leaderboards complicated our assessment.</p>\n<p>In the end, I made minor adjustments to the CSV file that contained the 98% accuracy results, which yielded a score of 1. I had a doubt because some entries I believed were AI-generated turned out to be real data, and vice versa. This uncertainty contributed to my doubts about the results.</p>\n<p>There was also an approach where we actually compared the test data with the training data and compared the percentages and made our inference based on that as well</p>",
      "votes": null,
      "replies": [
        {
          "id": 3103322,
          "author_name": "naweendissanayake",
          "author_url": "",
          "post_date": "01/23/2025 10:22:35",
          "content": "<p>Thank you very much for your clarification. Unfortunately, someone changed the submission file by hand; I do not see any computer vision there.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "3100987": "Hello,\n\nI'm curious to know about the techniques used by people who achieve 1.00. If possible, could you share these techniques to understand other people better?\n\nThank you.",
    "3101028": "The competition lasted for a long time, and there was no distinction between the public and private leaderboards. I only achieved a score of 1 through basic leaderboard testing, which didn't feel meaningful. After learning the answers, I conducted experiments, but I couldn't improve my score through the model. As time passed, my motivation decreased, and I stopped investing more time in enhancing the model's performance.\n\nOrganizing a good competition is difficult; it requires high-quality data and a strong organizing team. This might be the difference between community-driven competitions and those with cash prizes. Personally, I believe that as long as you’ve learned something, that’s enough. There are still many valuable lessons to be learned from open-source notebooks.",
    "3101039": "Thanks for your reply. I was little dissapointed about the compettion too.",
    "3103312": "We implemented multiple methods and models to achieve high accuracy in our project. Initially, we utilized models such as ViT (Vision Transformer) with configurations vit_base_patch16_224 and vit_large_patch16_224, along with ResNeXt101 and a Microsoft model. These efforts alone allowed us to reach approximately 93% accuracy.\n\nTo further enhance our results, we incorporated various preprocessing techniques. This step proved beneficial, boosting our accuracy to 98%. However, despite our best efforts, we could not improve beyond this threshold. The competition was prolonged, and the absence of a distinction between the public and private leaderboards complicated our assessment.\n\nIn the end, I made minor adjustments to the CSV file that contained the 98% accuracy results, which yielded a score of 1. I had a doubt because some entries I believed were AI-generated turned out to be real data, and vice versa. This uncertainty contributed to my doubts about the results.\n\nThere was also an approach where we actually compared the test data with the training data and compared the percentages and made our inference based on that as well",
    "3103322": "Thank you very much for your clarification. Unfortunately, someone changed the submission file by hand; I do not see any computer vision there."
  },
  "source": "meta"
}