{
  "id": 579727,
  "title": "Performance of trained models solely based on the competition data",
  "url": "/competitions/waveform-inversion/discussion/579727",
  "author_name": "",
  "post_date": "2025-05-20T01:51:29.423348100Z",
  "votes": 1,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Folks, </p>\n<p>What are the typical scores of models which are solely trained on the competition data? I am not talking about transfer learning of pre-trained models or, models trained on the larger dataset from OpenFWI. </p>",
  "messages": [
    {
      "id": "3205479",
      "postDate": "05/20/2025 01:51:29",
      "content": "<p>Folks, </p>\n<p>What are the typical scores of models which are solely trained on the competition data? I am not talking about transfer learning of pre-trained models or, models trained on the larger dataset from OpenFWI. </p>",
      "rawMarkdown": "Folks, \n\nWhat are the typical scores of models which are solely trained on the competition data? I am not talking about transfer learning of pre-trained models or, models trained on the larger dataset from OpenFWI.",
      "votes": null
    },
    {
      "id": "3205692",
      "postDate": "05/20/2025 08:37:47",
      "content": "<p>See me <a href=\"https://www.kaggle.com/competitions/waveform-inversion/discussion/574495\" target=\"_blank\">post</a>. It's low. If you can't train on the full dataset, don't bother to join. (Then again, it's not that hard to train on the full dataset, it's only 50GB or so after compression- see the public notebooks).</p>",
      "rawMarkdown": "See me [post](https://www.kaggle.com/competitions/waveform-inversion/discussion/574495). It's low. If you can't train on the full dataset, don't bother to join. (Then again, it's not that hard to train on the full dataset, it's only 50GB or so after compression- see the public notebooks).",
      "votes": null
    },
    {
      "id": "3205806",
      "postDate": "05/20/2025 11:29:28",
      "content": "<p>Thanks Greysnow. It is very helpful. With competition data, training from scratch with a radically different approach, I cannot get validation score to be below 200. I will use the full open Fwi Dataset. </p>",
      "rawMarkdown": "Thanks Greysnow. It is very helpful. With competition data, training from scratch with a radically different approach, I cannot get validation score to be below 200. I will use the full open Fwi Dataset.",
      "votes": null
    },
    {
      "id": "3205812",
      "postDate": "05/20/2025 11:42:03",
      "content": "<p>Yes, 200 is probably around the limit for only conpetition data.</p>",
      "rawMarkdown": "Yes, 200 is probably around the limit for only conpetition data.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3205692,
      "author_name": "shlomoron",
      "author_url": "",
      "post_date": "05/20/2025 08:37:47",
      "content": "<p>See me <a href=\"https://www.kaggle.com/competitions/waveform-inversion/discussion/574495\" target=\"_blank\">post</a>. It's low. If you can't train on the full dataset, don't bother to join. (Then again, it's not that hard to train on the full dataset, it's only 50GB or so after compression- see the public notebooks).</p>",
      "votes": null,
      "replies": [
        {
          "id": 3205806,
          "author_name": "sukantabasu",
          "author_url": "",
          "post_date": "05/20/2025 11:29:28",
          "content": "<p>Thanks Greysnow. It is very helpful. With competition data, training from scratch with a radically different approach, I cannot get validation score to be below 200. I will use the full open Fwi Dataset. </p>",
          "votes": null,
          "replies": [
            {
              "id": 3205812,
              "author_name": "shlomoron",
              "author_url": "",
              "post_date": "05/20/2025 11:42:03",
              "content": "<p>Yes, 200 is probably around the limit for only conpetition data.</p>",
              "votes": null,
              "replies": []
            }
          ]
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "3205479": "Folks, \n\nWhat are the typical scores of models which are solely trained on the competition data? I am not talking about transfer learning of pre-trained models or, models trained on the larger dataset from OpenFWI.",
    "3205692": "See me [post](https://www.kaggle.com/competitions/waveform-inversion/discussion/574495). It's low. If you can't train on the full dataset, don't bother to join. (Then again, it's not that hard to train on the full dataset, it's only 50GB or so after compression- see the public notebooks).",
    "3205806": "Thanks Greysnow. It is very helpful. With competition data, training from scratch with a radically different approach, I cannot get validation score to be below 200. I will use the full open Fwi Dataset.",
    "3205812": "Yes, 200 is probably around the limit for only conpetition data."
  },
  "source": "meta"
}