{
  "id": 591120,
  "title": "Does the test set include [orientation, behavior, phase]?",
  "url": "/competitions/cmi-detect-behavior-with-sensor-data/discussion/591120",
  "author_name": "",
  "post_date": "2025-07-25T14:27:53.551732600Z",
  "votes": 3,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Do these columns not exist in the official test.csv? Does this mean that they are not needed during training, or are these columns still useful? I am a beginner and would appreciate guidance from experts.</p>",
  "messages": [
    {
      "id": "3253942",
      "postDate": "07/25/2025 14:27:53",
      "content": "<p>Do these columns not exist in the official test.csv? Does this mean that they are not needed during training, or are these columns still useful? I am a beginner and would appreciate guidance from experts.</p>",
      "rawMarkdown": "Do these columns not exist in the official test.csv? Does this mean that they are not needed during training, or are these columns still useful? I am a beginner and would appreciate guidance from experts.",
      "votes": null
    },
    {
      "id": "3254048",
      "postDate": "07/25/2025 18:11:31",
      "content": "<p>Wow Kagglers are tough with the downvotes in this competition!</p>\n<p>As written in the <a href=\"https://www.kaggle.com/competitions/cmi-detect-behavior-with-sensor-data/data\" target=\"_blank\">data description</a> <code>gesture</code> and <code>orientation</code> are train only, meaning that they won't be given during the test phase. The <code>behavior</code> column I think is not accessible either, but this is not written in the data description which is a mistake I guess. The <code>phase</code> is not in the data dictionary, which is also a mistake.</p>\n<p>This means that you can't rely on them for your inference, but you can still use them somehow during training. For example you could try a stacking where a first model predicts  the current orientation and a second model uses that prediction to make the final answer. Not saying that this is actually useful in this competition, just that it is definitely possible!</p>",
      "rawMarkdown": "Wow Kagglers are tough with the downvotes in this competition!\n\nAs written in the [data description](https://www.kaggle.com/competitions/cmi-detect-behavior-with-sensor-data/data) ` gesture` and `orientation` are train only, meaning that they won't be given during the test phase. The `behavior` column I think is not accessible either, but this is not written in the data description which is a mistake I guess. The `phase` is not in the data dictionary, which is also a mistake.\n\nThis means that you can't rely on them for your inference, but you can still use them somehow during training. For example you could try a stacking where a first model predicts  the current orientation and a second model uses that prediction to make the final answer. Not saying that this is actually useful in this competition, just that it is definitely possible!",
      "votes": null
    },
    {
      "id": "3265112",
      "postDate": "08/07/2025 03:39:49",
      "content": "<p>Even I had the same doubt when I started, then finally decided to use only the IMU data for training to acheive a baseline score and then improve from there. <br>\nAs <a href=\"https://www.kaggle.com/optimo\" target=\"_blank\">@optimo</a>  says, for training the model its better NOT to rely on the columns that are NOT listed in the test.csv file.</p>",
      "rawMarkdown": "Even I had the same doubt when I started, then finally decided to use only the IMU data for training to acheive a baseline score and then improve from there. \nAs @optimo  says, for training the model its better NOT to rely on the columns that are NOT listed in the test.csv file.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3254048,
      "author_name": "optimo",
      "author_url": "",
      "post_date": "07/25/2025 18:11:31",
      "content": "<p>Wow Kagglers are tough with the downvotes in this competition!</p>\n<p>As written in the <a href=\"https://www.kaggle.com/competitions/cmi-detect-behavior-with-sensor-data/data\" target=\"_blank\">data description</a> <code>gesture</code> and <code>orientation</code> are train only, meaning that they won't be given during the test phase. The <code>behavior</code> column I think is not accessible either, but this is not written in the data description which is a mistake I guess. The <code>phase</code> is not in the data dictionary, which is also a mistake.</p>\n<p>This means that you can't rely on them for your inference, but you can still use them somehow during training. For example you could try a stacking where a first model predicts  the current orientation and a second model uses that prediction to make the final answer. Not saying that this is actually useful in this competition, just that it is definitely possible!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 3265112,
      "author_name": "gnanasenthilpro",
      "author_url": "",
      "post_date": "08/07/2025 03:39:49",
      "content": "<p>Even I had the same doubt when I started, then finally decided to use only the IMU data for training to acheive a baseline score and then improve from there. <br>\nAs <a href=\"https://www.kaggle.com/optimo\" target=\"_blank\">@optimo</a>  says, for training the model its better NOT to rely on the columns that are NOT listed in the test.csv file.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3253942": "Do these columns not exist in the official test.csv? Does this mean that they are not needed during training, or are these columns still useful? I am a beginner and would appreciate guidance from experts.",
    "3254048": "Wow Kagglers are tough with the downvotes in this competition!\n\nAs written in the [data description](https://www.kaggle.com/competitions/cmi-detect-behavior-with-sensor-data/data) ` gesture` and `orientation` are train only, meaning that they won't be given during the test phase. The `behavior` column I think is not accessible either, but this is not written in the data description which is a mistake I guess. The `phase` is not in the data dictionary, which is also a mistake.\n\nThis means that you can't rely on them for your inference, but you can still use them somehow during training. For example you could try a stacking where a first model predicts  the current orientation and a second model uses that prediction to make the final answer. Not saying that this is actually useful in this competition, just that it is definitely possible!",
    "3265112": "Even I had the same doubt when I started, then finally decided to use only the IMU data for training to acheive a baseline score and then improve from there. \nAs @optimo  says, for training the model its better NOT to rely on the columns that are NOT listed in the test.csv file."
  },
  "source": "meta"
}