{
  "id": 162063,
  "title": "LB score precision",
  "url": "/competitions/birdsong-recognition/discussion/162063",
  "author_name": "",
  "post_date": "2020-06-27T07:56:55.249445800Z",
  "votes": 20,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Could you increase the precision on LB by 1-2 additional digits?\nProper validation data would be even better :)</p>",
  "messages": [
    {
      "id": "903913",
      "postDate": "06/27/2020 07:56:55",
      "content": "<p>Could you increase the precision on LB by 1-2 additional digits?\nProper validation data would be even better :)</p>",
      "rawMarkdown": "Could you increase the precision on LB by 1-2 additional digits?\nProper validation data would be even better :)",
      "votes": null
    },
    {
      "id": "918602",
      "postDate": "07/07/2020 10:53:38",
      "content": "<p>Seconded!</p>\n\n<p>Unfortunately, the LB score is the best (only?) way we have of telling if we are increasing our solution's ability to generalize to domain shifted data... I am trying to use the provided soundscape recordings as a second validation set, but I am not sure how closely they resemble test set data.</p>\n\n<p>Having more precision could help us in understanding what works and what doesn't on addressing the core of the problem (domain adaptation). At the moment I am running into an issue - I would like to figure out which methods of scaling inputs generalizes better to domain shifted data, but two vastly different method achieve the same score (0.55 on LB). Being able to use the LB as signal into addressing such dilemmas could be very helpful to efforts towards addressing the core of the problem.</p>",
      "rawMarkdown": "Seconded!\n\nUnfortunately, the LB score is the best (only?) way we have of telling if we are increasing our solution's ability to generalize to domain shifted data... I am trying to use the provided soundscape recordings as a second validation set, but I am not sure how closely they resemble test set data.\n\nHaving more precision could help us in understanding what works and what doesn't on addressing the core of the problem (domain adaptation). At the moment I am running into an issue - I would like to figure out which methods of scaling inputs generalizes better to domain shifted data, but two vastly different method achieve the same score (0.55 on LB). Being able to use the LB as signal into addressing such dilemmas could be very helpful to efforts towards addressing the core of the problem.",
      "votes": null
    },
    {
      "id": "918935",
      "postDate": "07/07/2020 15:48:54",
      "content": "<p>I'll go ahead and bump it up by a digit now.</p>",
      "rawMarkdown": "I'll go ahead and bump it up by a digit now.",
      "votes": null
    },
    {
      "id": "919222",
      "postDate": "07/07/2020 19:04:30",
      "content": "<p>Thank you so much !</p>\n\n<p>This update motivates me more 🙌 </p>",
      "rawMarkdown": "Thank you so much !\n\nThis update motivates me more 🙌",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 918602,
      "author_name": "radek1",
      "author_url": "",
      "post_date": "07/07/2020 10:53:38",
      "content": "<p>Seconded!</p>\n\n<p>Unfortunately, the LB score is the best (only?) way we have of telling if we are increasing our solution's ability to generalize to domain shifted data... I am trying to use the provided soundscape recordings as a second validation set, but I am not sure how closely they resemble test set data.</p>\n\n<p>Having more precision could help us in understanding what works and what doesn't on addressing the core of the problem (domain adaptation). At the moment I am running into an issue - I would like to figure out which methods of scaling inputs generalizes better to domain shifted data, but two vastly different method achieve the same score (0.55 on LB). Being able to use the LB as signal into addressing such dilemmas could be very helpful to efforts towards addressing the core of the problem.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 918935,
      "author_name": "sohier",
      "author_url": "",
      "post_date": "07/07/2020 15:48:54",
      "content": "<p>I'll go ahead and bump it up by a digit now.</p>",
      "votes": null,
      "replies": [
        {
          "id": 919222,
          "author_name": "ttahara",
          "author_url": "",
          "post_date": "07/07/2020 19:04:30",
          "content": "<p>Thank you so much !</p>\n\n<p>This update motivates me more 🙌 </p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "903913": "Could you increase the precision on LB by 1-2 additional digits?\nProper validation data would be even better :)",
    "918602": "Seconded!\n\nUnfortunately, the LB score is the best (only?) way we have of telling if we are increasing our solution's ability to generalize to domain shifted data... I am trying to use the provided soundscape recordings as a second validation set, but I am not sure how closely they resemble test set data.\n\nHaving more precision could help us in understanding what works and what doesn't on addressing the core of the problem (domain adaptation). At the moment I am running into an issue - I would like to figure out which methods of scaling inputs generalizes better to domain shifted data, but two vastly different method achieve the same score (0.55 on LB). Being able to use the LB as signal into addressing such dilemmas could be very helpful to efforts towards addressing the core of the problem.",
    "918935": "I'll go ahead and bump it up by a digit now.",
    "919222": "Thank you so much !\n\nThis update motivates me more 🙌"
  },
  "source": "meta"
}