{
  "id": 664765,
  "title": "Is it overfitting?",
  "url": "/competitions/csiro-biomass/discussion/664765",
  "author_name": "",
  "post_date": "2025-12-27T17:02:05.378549900Z",
  "votes": 1,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Hi everyone,</p>\n<p>When I use the DINOv2 baseline code, my cv score is 0.659+, but after submitting, my leaderboard (LB) score is only 0.58. However, when I previously used ConvNeXt Tiny, my CV was 0.58–0.59 and my LB was also 0.58. Why is this happening? Has anyone encountered a similar situation?\nOf course, I used some data augmentation methods.</p>",
  "messages": [
    {
      "id": "3382492",
      "postDate": "12/27/2025 17:02:05",
      "content": "<p>Hi everyone,</p>\n<p>When I use the DINOv2 baseline code, my cv score is 0.659+, but after submitting, my leaderboard (LB) score is only 0.58. However, when I previously used ConvNeXt Tiny, my CV was 0.58–0.59 and my LB was also 0.58. Why is this happening? Has anyone encountered a similar situation?\nOf course, I used some data augmentation methods.</p>",
      "rawMarkdown": "Hi everyone,\n\nWhen I use the DINOv2 baseline code, my cv score is 0.659+, but after submitting, my leaderboard (LB) score is only 0.58. However, when I previously used ConvNeXt Tiny, my CV was 0.58–0.59 and my LB was also 0.58. Why is this happening? Has anyone encountered a similar situation?\nOf course, I used some data augmentation methods.",
      "votes": null
    },
    {
      "id": "3382501",
      "postDate": "12/27/2025 17:34:33",
      "content": "<p>Yes, overfitting(cv: 73  &gt; Lb: 65; since there is little data, the model has too much capacity. </p>",
      "rawMarkdown": "Yes, overfitting(cv: 73  > Lb: 65; since there is little data, the model has too much capacity.",
      "votes": null
    },
    {
      "id": "3382503",
      "postDate": "12/27/2025 17:50:34",
      "content": "<p>This discussion is helpful: <a href=\"https://www.kaggle.com/competitions/csiro-biomass/discussion/615401\" target=\"_blank\">Splitting Discussion</a></p>\n<p>I have found, as have others, that some splits start to fall apart as you improve on them. So you'll have a nice correlation up to a point.</p>\n<p>I've started to use different splits for different purposes because of this. Some are to understand the problem, others are to compare models, etc.</p>",
      "rawMarkdown": "This discussion is helpful: [Splitting Discussion](https://www.kaggle.com/competitions/csiro-biomass/discussion/615401)\n\nI have found, as have others, that some splits start to fall apart as you improve on them. So you'll have a nice correlation up to a point.\n\nI've started to use different splits for different purposes because of this. Some are to understand the problem, others are to compare models, etc.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3382501,
      "author_name": "giovannyrodrguez",
      "author_url": "",
      "post_date": "12/27/2025 17:34:33",
      "content": "<p>Yes, overfitting(cv: 73  &gt; Lb: 65; since there is little data, the model has too much capacity. </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 3382503,
      "author_name": "chemdatafarmer",
      "author_url": "",
      "post_date": "12/27/2025 17:50:34",
      "content": "<p>This discussion is helpful: <a href=\"https://www.kaggle.com/competitions/csiro-biomass/discussion/615401\" target=\"_blank\">Splitting Discussion</a></p>\n<p>I have found, as have others, that some splits start to fall apart as you improve on them. So you'll have a nice correlation up to a point.</p>\n<p>I've started to use different splits for different purposes because of this. Some are to understand the problem, others are to compare models, etc.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3382492": "Hi everyone,\n\nWhen I use the DINOv2 baseline code, my cv score is 0.659+, but after submitting, my leaderboard (LB) score is only 0.58. However, when I previously used ConvNeXt Tiny, my CV was 0.58–0.59 and my LB was also 0.58. Why is this happening? Has anyone encountered a similar situation?\nOf course, I used some data augmentation methods.",
    "3382501": "Yes, overfitting(cv: 73  > Lb: 65; since there is little data, the model has too much capacity.",
    "3382503": "This discussion is helpful: [Splitting Discussion](https://www.kaggle.com/competitions/csiro-biomass/discussion/615401)\n\nI have found, as have others, that some splits start to fall apart as you improve on them. So you'll have a nice correlation up to a point.\n\nI've started to use different splits for different purposes because of this. Some are to understand the problem, others are to compare models, etc."
  },
  "source": "meta"
}