{
  "id": 514842,
  "title": "Public LB : How much representative is it? CV vs LB",
  "url": "/competitions/leash-BELKA/discussion/514842",
  "author_name": "AC",
  "post_date": "2024-06-25T18:13:45.917000",
  "votes": 3,
  "comment_count": 6,
  "views": 0,
  "content": "<p>Given, 2/3rd of the molecules are not in the training distribution, how much is the public lb representative of the final private lb. Should we solely rely on public lb or cv scores? <br>\nWith the addition of new library, how much of a difference it can make? </p>",
  "messages": [
    {
      "id": 2889881,
      "postDate": "2024-06-25T18:13:45.917Z",
      "content": "<p>Given, 2/3rd of the molecules are not in the training distribution, how much is the public lb representative of the final private lb. Should we solely rely on public lb or cv scores? <br>\nWith the addition of new library, how much of a difference it can make? </p>",
      "rawMarkdown": "Given, 2/3rd of the molecules are not in the training distribution, how much is the public lb representative of the final private lb. Should we solely rely on public lb or cv scores? \nWith the addition of new library, how much of a difference it can make? ",
      "votes": 3
    },
    {
      "id": 2893415,
      "postDate": "2024-06-27T19:34:10.360Z",
      "content": "<p>I sense it is going to be a big shuffle</p>",
      "rawMarkdown": "I sense it is going to be a big shuffle",
      "votes": 4,
      "replies": [
        {
          "id": 2898662,
          "postDate": "2024-07-01T09:10:32.500Z",
          "content": "<p>When I sample 10M data, LB score always better than using all data…</p>",
          "rawMarkdown": "When I sample 10M data, LB score always better than using all data...",
          "replies": [
            {
              "id": 2898725,
              "postDate": "2024-07-01T09:54:01.603Z",
              "content": "<p>With the same models, same hyper-parameters? </p>",
              "rawMarkdown": "With the same models, same hyper-parameters? "
            },
            {
              "id": 2899573,
              "postDate": "2024-07-01T18:18:08.443Z",
              "rawMarkdown": "",
              "isDeleted": true
            },
            {
              "id": 2899575,
              "postDate": "2024-07-01T18:18:51.897Z",
              "content": "<p>We tried same model type but bigger lr &amp; bs</p>",
              "rawMarkdown": "We tried same model type but bigger lr & bs",
              "votes": 1
            },
            {
              "id": 2901007,
              "postDate": "2024-07-02T15:49:47.783Z",
              "content": "<p>Thank you for the response! :)</p>",
              "rawMarkdown": "Thank you for the response! :)"
            }
          ]
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 2893415,
      "author_name": "yamu_duck",
      "author_url": "",
      "post_date": "2024-06-27T19:34:10.360000",
      "content": "<p>I sense it is going to be a big shuffle</p>",
      "votes": 4,
      "replies": [
        {
          "id": 2898662,
          "author_name": "go",
          "author_url": "",
          "post_date": "2024-07-01T09:10:32.500000",
          "content": "<p>When I sample 10M data, LB score always better than using all data…</p>",
          "votes": 0,
          "replies": [
            {
              "id": 2898725,
              "author_name": "AC",
              "author_url": "",
              "post_date": "2024-07-01T09:54:01.603000",
              "content": "<p>With the same models, same hyper-parameters? </p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 2899573,
              "author_name": "",
              "author_url": "",
              "post_date": "2024-07-01T18:18:08.443000",
              "content": "",
              "votes": 0,
              "replies": []
            },
            {
              "id": 2899575,
              "author_name": "Angle",
              "author_url": "",
              "post_date": "2024-07-01T18:18:51.897000",
              "content": "<p>We tried same model type but bigger lr &amp; bs</p>",
              "votes": 1,
              "replies": []
            },
            {
              "id": 2901007,
              "author_name": "AC",
              "author_url": "",
              "post_date": "2024-07-02T15:49:47.783000",
              "content": "<p>Thank you for the response! :)</p>",
              "votes": 0,
              "replies": []
            }
          ]
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "2889881": "Given, 2/3rd of the molecules are not in the training distribution, how much is the public lb representative of the final private lb. Should we solely rely on public lb or cv scores? \nWith the addition of new library, how much of a difference it can make? ",
    "2893415": "I sense it is going to be a big shuffle"
  }
}