{
  "id": 664542,
  "title": "How good is your score compared to all authentic baseline?",
  "url": "/competitions/recodai-luc-scientific-image-forgery-detection/discussion/664542",
  "author_name": "",
  "post_date": "2025-12-26T05:53:54.841418500Z",
  "votes": 6,
  "comment_count": 6,
  "views": 0,
  "content": "<p>I'm finally getting some results with my FPN model. Here is the metrics I'm tracking.</p>\n<p>Baseline score is 0.4592. TP contribution is +0.008 and FP contribution is -0.004, so the overall improvement is 0.004. What's your score look like? I wanna see the ceiling of this problem.</p>\n<pre><code>  \"total_samples\": 5176,\n  \"accuracy\": 0.499806800618238,\n  \"precision\": 0.9166666666666666,\n  \"recall\": 0.08252947481243302,\n  \"count_forged_correct (TP)\": 231,\n  \"count_authentic_correct (TN)\": 2356,\n  \"count_false_positives (FP)\": 21,\n  \"count_missed_forgeries (FN)\": 2568,\n  \"tp_impact\": 0.008754780732259536,\n  \"fp_impact\": 0.004057187017001545,\n  \"mean_image_score\": 0.4639325241634806,\n  \"baseline_score\": 0.45923493044822256,\n  \"lift_over_baseline\": 0.004697593715258019,\n</code></pre>",
  "messages": [
    {
      "id": "3381981",
      "postDate": "12/26/2025 05:53:54",
      "content": "<p>I'm finally getting some results with my FPN model. Here is the metrics I'm tracking.</p>\n<p>Baseline score is 0.4592. TP contribution is +0.008 and FP contribution is -0.004, so the overall improvement is 0.004. What's your score look like? I wanna see the ceiling of this problem.</p>\n<pre><code>  \"total_samples\": 5176,\n  \"accuracy\": 0.499806800618238,\n  \"precision\": 0.9166666666666666,\n  \"recall\": 0.08252947481243302,\n  \"count_forged_correct (TP)\": 231,\n  \"count_authentic_correct (TN)\": 2356,\n  \"count_false_positives (FP)\": 21,\n  \"count_missed_forgeries (FN)\": 2568,\n  \"tp_impact\": 0.008754780732259536,\n  \"fp_impact\": 0.004057187017001545,\n  \"mean_image_score\": 0.4639325241634806,\n  \"baseline_score\": 0.45923493044822256,\n  \"lift_over_baseline\": 0.004697593715258019,\n</code></pre>",
      "rawMarkdown": "I'm finally getting some results with my FPN model. Here is the metrics I'm tracking.\n\nBaseline score is 0.4592. TP contribution is +0.008 and FP contribution is -0.004, so the overall improvement is 0.004. What's your score look like? I wanna see the ceiling of this problem.\n\n```python\n  \"total_samples\": 5176,\n  \"accuracy\": 0.499806800618238,\n  \"precision\": 0.9166666666666666,\n  \"recall\": 0.08252947481243302,\n  \"count_forged_correct (TP)\": 231,\n  \"count_authentic_correct (TN)\": 2356,\n  \"count_false_positives (FP)\": 21,\n  \"count_missed_forgeries (FN)\": 2568,\n  \"tp_impact\": 0.008754780732259536,\n  \"fp_impact\": 0.004057187017001545,\n  \"mean_image_score\": 0.4639325241634806,\n  \"baseline_score\": 0.45923493044822256,\n  \"lift_over_baseline\": 0.004697593715258019,\n```",
      "votes": null
    },
    {
      "id": "3382268",
      "postDate": "12/27/2025 04:48:36",
      "content": "<p>mine is 0.4763 with LB score 0.303</p>",
      "rawMarkdown": "mine is 0.4763 with LB score 0.303",
      "votes": null
    },
    {
      "id": "3382345",
      "postDate": "12/27/2025 09:17:02",
      "content": "<p>I think we should also track our predictions on supplemental data. My model always predicts no mask for those samples so they became all authentic. I think same thing happens on the public test set.</p>",
      "rawMarkdown": "I think we should also track our predictions on supplemental data. My model always predicts no mask for those samples so they became all authentic. I think same thing happens on the public test set.",
      "votes": null
    },
    {
      "id": "3382352",
      "postDate": "12/27/2025 09:30:50",
      "content": "<p>yeah, I haven't used supplemental data properly yet. But so far it is pretty rough to train on the main images</p>",
      "rawMarkdown": "yeah, I haven't used supplemental data properly yet. But so far it is pretty rough to train on the main images",
      "votes": null
    },
    {
      "id": "3382638",
      "postDate": "12/28/2025 06:24:10",
      "content": "<p>update: i have managed to push CV score to 0.5183. F1 on Authentic samples: 0.98, F1 on Forged samples: 0.12</p>\n<p>LB score 0.303. </p>\n<p>Lower threshhold to 0.33 make LB score 0.201</p>",
      "rawMarkdown": "update: i have managed to push CV score to 0.5183. F1 on Authentic samples: 0.98, F1 on Forged samples: 0.12\n\nLB score 0.303. \n\nLower threshhold to 0.33 make LB score 0.201",
      "votes": null
    },
    {
      "id": "3382642",
      "postDate": "12/28/2025 06:37:17",
      "content": "<p>Nice improvement, can I ask what your model is?</p>",
      "rawMarkdown": "Nice improvement, can I ask what your model is?",
      "votes": null
    },
    {
      "id": "3382701",
      "postDate": "12/28/2025 10:05:32",
      "content": "<p>i copied Segmenter model from public kernel and adding classification head</p>",
      "rawMarkdown": "i copied Segmenter model from public kernel and adding classification head",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3382268,
      "author_name": "llkh0a",
      "author_url": "",
      "post_date": "12/27/2025 04:48:36",
      "content": "<p>mine is 0.4763 with LB score 0.303</p>",
      "votes": null,
      "replies": [
        {
          "id": 3382345,
          "author_name": "gunesevitan",
          "author_url": "",
          "post_date": "12/27/2025 09:17:02",
          "content": "<p>I think we should also track our predictions on supplemental data. My model always predicts no mask for those samples so they became all authentic. I think same thing happens on the public test set.</p>",
          "votes": null,
          "replies": [
            {
              "id": 3382638,
              "author_name": "llkh0a",
              "author_url": "",
              "post_date": "12/28/2025 06:24:10",
              "content": "<p>update: i have managed to push CV score to 0.5183. F1 on Authentic samples: 0.98, F1 on Forged samples: 0.12</p>\n<p>LB score 0.303. </p>\n<p>Lower threshhold to 0.33 make LB score 0.201</p>",
              "votes": null,
              "replies": [
                {
                  "id": 3382642,
                  "author_name": "gunesevitan",
                  "author_url": "",
                  "post_date": "12/28/2025 06:37:17",
                  "content": "<p>Nice improvement, can I ask what your model is?</p>",
                  "votes": null,
                  "replies": [
                    {
                      "id": 3382701,
                      "author_name": "llkh0a",
                      "author_url": "",
                      "post_date": "12/28/2025 10:05:32",
                      "content": "<p>i copied Segmenter model from public kernel and adding classification head</p>",
                      "votes": null,
                      "replies": []
                    }
                  ]
                }
              ]
            }
          ]
        },
        {
          "id": 3382352,
          "author_name": "llkh0a",
          "author_url": "",
          "post_date": "12/27/2025 09:30:50",
          "content": "<p>yeah, I haven't used supplemental data properly yet. But so far it is pretty rough to train on the main images</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "3381981": "I'm finally getting some results with my FPN model. Here is the metrics I'm tracking.\n\nBaseline score is 0.4592. TP contribution is +0.008 and FP contribution is -0.004, so the overall improvement is 0.004. What's your score look like? I wanna see the ceiling of this problem.\n\n```python\n  \"total_samples\": 5176,\n  \"accuracy\": 0.499806800618238,\n  \"precision\": 0.9166666666666666,\n  \"recall\": 0.08252947481243302,\n  \"count_forged_correct (TP)\": 231,\n  \"count_authentic_correct (TN)\": 2356,\n  \"count_false_positives (FP)\": 21,\n  \"count_missed_forgeries (FN)\": 2568,\n  \"tp_impact\": 0.008754780732259536,\n  \"fp_impact\": 0.004057187017001545,\n  \"mean_image_score\": 0.4639325241634806,\n  \"baseline_score\": 0.45923493044822256,\n  \"lift_over_baseline\": 0.004697593715258019,\n```",
    "3382268": "mine is 0.4763 with LB score 0.303",
    "3382345": "I think we should also track our predictions on supplemental data. My model always predicts no mask for those samples so they became all authentic. I think same thing happens on the public test set.",
    "3382352": "yeah, I haven't used supplemental data properly yet. But so far it is pretty rough to train on the main images",
    "3382638": "update: i have managed to push CV score to 0.5183. F1 on Authentic samples: 0.98, F1 on Forged samples: 0.12\n\nLB score 0.303. \n\nLower threshhold to 0.33 make LB score 0.201",
    "3382642": "Nice improvement, can I ask what your model is?",
    "3382701": "i copied Segmenter model from public kernel and adding classification head"
  },
  "source": "meta"
}