{
  "id": 588122,
  "title": "Whether Pseudo-label work for you?",
  "url": "/competitions/cmi-detect-behavior-with-sensor-data/discussion/588122",
  "author_name": "",
  "post_date": "2025-07-04T14:27:54.838163300Z",
  "votes": 6,
  "comment_count": 5,
  "views": 0,
  "content": "<p>Because of the test file is much more smaller than train file, so I'm not sure whether it worth to do pseudo-label. But I still want to know whether someone tried and whether it work even if little improvement?</p>",
  "messages": [
    {
      "id": "3241091",
      "postDate": "07/04/2025 14:27:54",
      "content": "<p>Because of the test file is much more smaller than train file, so I'm not sure whether it worth to do pseudo-label. But I still want to know whether someone tried and whether it work even if little improvement?</p>",
      "rawMarkdown": "Because of the test file is much more smaller than train file, so I'm not sure whether it worth to do pseudo-label. But I still want to know whether someone tried and whether it work even if little improvement?",
      "votes": null
    },
    {
      "id": "3241984",
      "postDate": "07/05/2025 12:14:58",
      "content": "<p>I haven't done pseudo labelling before, but I think you mean creating labels for the two test sequences so that we can include them in training models? </p>\n<p>Because there are only two sequences, it would be interesting to have community discussion as to what the correct labels are.  It does seem like the payoff should be small, but maybe it's worth doing if you have high confidence in the labels.</p>",
      "rawMarkdown": "I haven't done pseudo labelling before, but I think you mean creating labels for the two test sequences so that we can include them in training models? \n\nBecause there are only two sequences, it would be interesting to have community discussion as to what the correct labels are.  It does seem like the payoff should be small, but maybe it's worth doing if you have high confidence in the labels.",
      "votes": null
    },
    {
      "id": "3242096",
      "postDate": "07/05/2025 14:56:17",
      "content": "<p>not sure if you would have enough time to train a model for each sequence</p>",
      "rawMarkdown": "not sure if you would have enough time to train a model for each sequence",
      "votes": null
    },
    {
      "id": "3242358",
      "postDate": "07/05/2025 19:50:18",
      "content": "<p>I've seen pseudo-labeling work in some cases, even with small test sets. It might not give huge gains, but every little bit counts. <a href=\"https://www.kaggle.com/kurisew\" target=\"_blank\">@kurisew</a> </p>",
      "rawMarkdown": "I've seen pseudo-labeling work in some cases, even with small test sets. It might not give huge gains, but every little bit counts. @kurisew",
      "votes": null
    },
    {
      "id": "3242414",
      "postDate": "07/05/2025 20:52:27",
      "content": "<p>Interesting work! I've found pseudo-labeling can sometimes help, even with limited test data especially when combined with strong base models. It might not boost performance dramatically, but small gains can make a difference. <a href=\"https://www.kaggle.com/kurisew\" target=\"_blank\">@kurisew</a></p>",
      "rawMarkdown": "Interesting work! I've found pseudo-labeling can sometimes help, even with limited test data especially when combined with strong base models. It might not boost performance dramatically, but small gains can make a difference. @kurisew",
      "votes": null
    },
    {
      "id": "3243454",
      "postDate": "07/07/2025 06:52:15",
      "content": "<p>This means your trained model to predict the test file then push them into training loop instead of &gt;train a model for each sequence</p>",
      "rawMarkdown": "This means your trained model to predict the test file then push them into training loop instead of >train a model for each sequence",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3241984,
      "author_name": "petermoorhouse",
      "author_url": "",
      "post_date": "07/05/2025 12:14:58",
      "content": "<p>I haven't done pseudo labelling before, but I think you mean creating labels for the two test sequences so that we can include them in training models? </p>\n<p>Because there are only two sequences, it would be interesting to have community discussion as to what the correct labels are.  It does seem like the payoff should be small, but maybe it's worth doing if you have high confidence in the labels.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 3242096,
      "author_name": "tahaalshatiri",
      "author_url": "",
      "post_date": "07/05/2025 14:56:17",
      "content": "<p>not sure if you would have enough time to train a model for each sequence</p>",
      "votes": null,
      "replies": [
        {
          "id": 3243454,
          "author_name": "kurisew",
          "author_url": "",
          "post_date": "07/07/2025 06:52:15",
          "content": "<p>This means your trained model to predict the test file then push them into training loop instead of &gt;train a model for each sequence</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 3242358,
      "author_name": "sadiashahidlatif",
      "author_url": "",
      "post_date": "07/05/2025 19:50:18",
      "content": "<p>I've seen pseudo-labeling work in some cases, even with small test sets. It might not give huge gains, but every little bit counts. <a href=\"https://www.kaggle.com/kurisew\" target=\"_blank\">@kurisew</a> </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 3242414,
      "author_name": "hijabzahra8",
      "author_url": "",
      "post_date": "07/05/2025 20:52:27",
      "content": "<p>Interesting work! I've found pseudo-labeling can sometimes help, even with limited test data especially when combined with strong base models. It might not boost performance dramatically, but small gains can make a difference. <a href=\"https://www.kaggle.com/kurisew\" target=\"_blank\">@kurisew</a></p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3241091": "Because of the test file is much more smaller than train file, so I'm not sure whether it worth to do pseudo-label. But I still want to know whether someone tried and whether it work even if little improvement?",
    "3241984": "I haven't done pseudo labelling before, but I think you mean creating labels for the two test sequences so that we can include them in training models? \n\nBecause there are only two sequences, it would be interesting to have community discussion as to what the correct labels are.  It does seem like the payoff should be small, but maybe it's worth doing if you have high confidence in the labels.",
    "3242096": "not sure if you would have enough time to train a model for each sequence",
    "3242358": "I've seen pseudo-labeling work in some cases, even with small test sets. It might not give huge gains, but every little bit counts. @kurisew",
    "3242414": "Interesting work! I've found pseudo-labeling can sometimes help, even with limited test data especially when combined with strong base models. It might not boost performance dramatically, but small gains can make a difference. @kurisew",
    "3243454": "This means your trained model to predict the test file then push them into training loop instead of >train a model for each sequence"
  },
  "source": "meta"
}