{
  "id": 665886,
  "title": "cv vs lb gap issues",
  "url": "/competitions/deep-past-initiative-machine-translation/discussion/665886",
  "author_name": "",
  "post_date": "2026-01-04T09:48:02.211714300Z",
  "votes": 9,
  "comment_count": 8,
  "views": 0,
  "content": "<p>I did not see many teams talked about CV vs LB, my cv is 47.6 (1 /5 fold) but lb only got 28.6. May I know if you guys have the same issues  as me.</p>",
  "messages": [
    {
      "id": "3385949",
      "postDate": "01/04/2026 09:48:02",
      "content": "<p>I did not see many teams talked about CV vs LB, my cv is 47.6 (1 /5 fold) but lb only got 28.6. May I know if you guys have the same issues  as me.</p>",
      "rawMarkdown": "I did not see many teams talked about CV vs LB, my cv is 47.6 (1 /5 fold) but lb only got 28.6. May I know if you guys have the same issues  as me.",
      "votes": null
    },
    {
      "id": "3385958",
      "postDate": "01/04/2026 09:58:20",
      "content": "<p>Some of us talked about that actually, one reason is the normalization of gaps (x, xxx to <code>&lt;gap&gt;</code> and <code>&lt;big_gap&gt;</code> <strong>both for transliterations and translations</strong>) But many of us still has a large margin between CV and LB. My CV also is 45-48 and LB varies between 32-33.7. It looks like we still have some normalization differences like gap normalization that hasn't been pointed out yet.</p>",
      "rawMarkdown": "Some of us talked about that actually, one reason is the normalization of gaps (x, xxx to `<gap>` and `<big_gap>` **both for transliterations and translations**) But many of us still has a large margin between CV and LB. My CV also is 45-48 and LB varies between 32-33.7. It looks like we still have some normalization differences like gap normalization that hasn't been pointed out yet.",
      "votes": null
    },
    {
      "id": "3385961",
      "postDate": "01/04/2026 10:00:26",
      "content": "<p>thanks so much for your reply </p>",
      "rawMarkdown": "thanks so much for your reply",
      "votes": null
    },
    {
      "id": "3392228",
      "postDate": "01/16/2026 13:59:22",
      "content": "<p>My CV is now around 55-56. Similar for you or no? </p>",
      "rawMarkdown": "My CV is now around 55-56. Similar for you or no?",
      "votes": null
    },
    {
      "id": "3392284",
      "postDate": "01/16/2026 16:02:04",
      "content": "<p>better than me, i can only got 45-46</p>",
      "rawMarkdown": "better than me, i can only got 45-46",
      "votes": null
    },
    {
      "id": "3400147",
      "postDate": "02/01/2026 02:34:40",
      "content": "<p>ahh interesting I need to train will see how much I get :) </p>",
      "rawMarkdown": "ahh interesting I need to train will see how much I get :)",
      "votes": null
    },
    {
      "id": "3400884",
      "postDate": "02/02/2026 14:39:26",
      "content": "<p>Running so far seems aligned but have to train more for higher scores (to be updated)</p>\n<ul>\n<li>CV 17.33 LB 20.9</li>\n<li>CV 17.57 LB 22.59</li>\n</ul>",
      "rawMarkdown": "Running so far seems aligned but have to train more for higher scores (to be updated)\n- CV 17.33 LB 20.9\n- CV 17.57 LB 22.59",
      "votes": null
    },
    {
      "id": "3400914",
      "postDate": "02/02/2026 15:42:07",
      "content": "<p>Hi! Do you mind giving a bit information of how you did it? I think most people use the validation scheme from the public kernels, which is 1 fold, and using that gives very high local validation score</p>",
      "rawMarkdown": "Hi! Do you mind giving a bit information of how you did it? I think most people use the validation scheme from the public kernels, which is 1 fold, and using that gives very high local validation score",
      "votes": null
    },
    {
      "id": "3401000",
      "postDate": "02/02/2026 18:19:30",
      "content": "<p>sure need to run more experiments to deduce correclty . will post</p>",
      "rawMarkdown": "sure need to run more experiments to deduce correclty . will post",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3385958,
      "author_name": "nlztrk",
      "author_url": "",
      "post_date": "01/04/2026 09:58:20",
      "content": "<p>Some of us talked about that actually, one reason is the normalization of gaps (x, xxx to <code>&lt;gap&gt;</code> and <code>&lt;big_gap&gt;</code> <strong>both for transliterations and translations</strong>) But many of us still has a large margin between CV and LB. My CV also is 45-48 and LB varies between 32-33.7. It looks like we still have some normalization differences like gap normalization that hasn't been pointed out yet.</p>",
      "votes": null,
      "replies": [
        {
          "id": 3385961,
          "author_name": "leehann",
          "author_url": "",
          "post_date": "01/04/2026 10:00:26",
          "content": "<p>thanks so much for your reply </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 3392228,
      "author_name": "jackvd",
      "author_url": "",
      "post_date": "01/16/2026 13:59:22",
      "content": "<p>My CV is now around 55-56. Similar for you or no? </p>",
      "votes": null,
      "replies": [
        {
          "id": 3392284,
          "author_name": "leehann",
          "author_url": "",
          "post_date": "01/16/2026 16:02:04",
          "content": "<p>better than me, i can only got 45-46</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 3400147,
      "author_name": "gauravbrills",
      "author_url": "",
      "post_date": "02/01/2026 02:34:40",
      "content": "<p>ahh interesting I need to train will see how much I get :) </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 3400884,
      "author_name": "gauravbrills",
      "author_url": "",
      "post_date": "02/02/2026 14:39:26",
      "content": "<p>Running so far seems aligned but have to train more for higher scores (to be updated)</p>\n<ul>\n<li>CV 17.33 LB 20.9</li>\n<li>CV 17.57 LB 22.59</li>\n</ul>",
      "votes": null,
      "replies": [
        {
          "id": 3400914,
          "author_name": "honganzhu",
          "author_url": "",
          "post_date": "02/02/2026 15:42:07",
          "content": "<p>Hi! Do you mind giving a bit information of how you did it? I think most people use the validation scheme from the public kernels, which is 1 fold, and using that gives very high local validation score</p>",
          "votes": null,
          "replies": [
            {
              "id": 3401000,
              "author_name": "gauravbrills",
              "author_url": "",
              "post_date": "02/02/2026 18:19:30",
              "content": "<p>sure need to run more experiments to deduce correclty . will post</p>",
              "votes": null,
              "replies": []
            }
          ]
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "3385949": "I did not see many teams talked about CV vs LB, my cv is 47.6 (1 /5 fold) but lb only got 28.6. May I know if you guys have the same issues  as me.",
    "3385958": "Some of us talked about that actually, one reason is the normalization of gaps (x, xxx to `<gap>` and `<big_gap>` **both for transliterations and translations**) But many of us still has a large margin between CV and LB. My CV also is 45-48 and LB varies between 32-33.7. It looks like we still have some normalization differences like gap normalization that hasn't been pointed out yet.",
    "3385961": "thanks so much for your reply",
    "3392228": "My CV is now around 55-56. Similar for you or no?",
    "3392284": "better than me, i can only got 45-46",
    "3400147": "ahh interesting I need to train will see how much I get :)",
    "3400884": "Running so far seems aligned but have to train more for higher scores (to be updated)\n- CV 17.33 LB 20.9\n- CV 17.57 LB 22.59",
    "3400914": "Hi! Do you mind giving a bit information of how you did it? I think most people use the validation scheme from the public kernels, which is 1 fold, and using that gives very high local validation score",
    "3401000": "sure need to run more experiments to deduce correclty . will post"
  },
  "source": "meta"
}