{
  "id": 665250,
  "title": "Discrepancy between Validation Loss and Metric when increasing Batch Size",
  "url": "/competitions/deep-past-initiative-machine-translation/discussion/665250",
  "author_name": "",
  "post_date": "2025-12-31T00:29:31.126542400Z",
  "votes": null,
  "comment_count": 3,
  "views": 0,
  "content": "<p>When I increase the batch size(from 2 to 32), I'm observing a strange divergence between the loss and the evaluation metrics. Specifically, the validation loss spikes dramatically, yet the competition evaluation metric continues to improve. Has anyone else encountered a similar situation?\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F8060745%2Ffeeeddd2ffa7c8304c18aaad7d0855d7%2F1.png?generation=1767140945545450&amp;alt=media\" alt=\"\"></p>",
  "messages": [
    {
      "id": "3383857",
      "postDate": "12/31/2025 00:29:31",
      "content": "<p>When I increase the batch size(from 2 to 32), I'm observing a strange divergence between the loss and the evaluation metrics. Specifically, the validation loss spikes dramatically, yet the competition evaluation metric continues to improve. Has anyone else encountered a similar situation?\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F8060745%2Ffeeeddd2ffa7c8304c18aaad7d0855d7%2F1.png?generation=1767140945545450&amp;alt=media\" alt=\"\"></p>",
      "rawMarkdown": "When I increase the batch size(from 2 to 32), I'm observing a strange divergence between the loss and the evaluation metrics. Specifically, the validation loss spikes dramatically, yet the competition evaluation metric continues to improve. Has anyone else encountered a similar situation?\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F8060745%2Ffeeeddd2ffa7c8304c18aaad7d0855d7%2F1.png?generation=1767140945545450&alt=media)",
      "votes": null
    },
    {
      "id": "3384042",
      "postDate": "12/31/2025 11:10:34",
      "content": "<p>Yup, my chrf dipped. </p>\n<p>Lesser data -&gt; More batch size -&gt; Lesser backprops -&gt; Less learning</p>",
      "rawMarkdown": "Yup, my chrf dipped. \n\n Lesser data -> More batch size -> Lesser backprops -> Less learning",
      "votes": null
    },
    {
      "id": "3384321",
      "postDate": "01/01/2026 01:37:55",
      "content": "<p>Does your loss also go down first and then go up?</p>",
      "rawMarkdown": "Does your loss also go down first and then go up?",
      "votes": null
    },
    {
      "id": "3384326",
      "postDate": "01/01/2026 01:50:12",
      "content": "<p>Eval loss? Yes </p>",
      "rawMarkdown": "Eval loss? Yes",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3384042,
      "author_name": "",
      "author_url": "",
      "post_date": "12/31/2025 11:10:34",
      "content": "<p>Yup, my chrf dipped. </p>\n<p>Lesser data -&gt; More batch size -&gt; Lesser backprops -&gt; Less learning</p>",
      "votes": null,
      "replies": [
        {
          "id": 3384321,
          "author_name": "act18l",
          "author_url": "",
          "post_date": "01/01/2026 01:37:55",
          "content": "<p>Does your loss also go down first and then go up?</p>",
          "votes": null,
          "replies": [
            {
              "id": 3384326,
              "author_name": "",
              "author_url": "",
              "post_date": "01/01/2026 01:50:12",
              "content": "<p>Eval loss? Yes </p>",
              "votes": null,
              "replies": []
            }
          ]
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "3383857": "When I increase the batch size(from 2 to 32), I'm observing a strange divergence between the loss and the evaluation metrics. Specifically, the validation loss spikes dramatically, yet the competition evaluation metric continues to improve. Has anyone else encountered a similar situation?\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F8060745%2Ffeeeddd2ffa7c8304c18aaad7d0855d7%2F1.png?generation=1767140945545450&alt=media)",
    "3384042": "Yup, my chrf dipped. \n\n Lesser data -> More batch size -> Lesser backprops -> Less learning",
    "3384321": "Does your loss also go down first and then go up?",
    "3384326": "Eval loss? Yes"
  },
  "source": "meta"
}