{
  "id": 679714,
  "title": "Low LB score",
  "url": "/competitions/deep-past-initiative-machine-translation/discussion/679714",
  "author_name": "",
  "post_date": "2026-03-03T11:26:44.181079900Z",
  "votes": -1,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Hello. </p>\n<p>I Preprocessed data - using  where required; Removed !, ? signs where required; removed also [ ], ( ); I have been training model for 12 epochs; Test and val loss decreases for all epochs. But LB score is around 10.</p>\n<p>I could imagen: splitting examples in sentences would be helpful, have not done it yet.</p>\n<p>My suspicion is: such a low score is an indication that pipeline is broken and data preprocessing is not a solution. Could someone share any experience/opinion that might be helpful?</p>",
  "messages": [
    {
      "id": "3416608",
      "postDate": "03/03/2026 11:26:44",
      "content": "<p>Hello. </p>\n<p>I Preprocessed data - using  where required; Removed !, ? signs where required; removed also [ ], ( ); I have been training model for 12 epochs; Test and val loss decreases for all epochs. But LB score is around 10.</p>\n<p>I could imagen: splitting examples in sentences would be helpful, have not done it yet.</p>\n<p>My suspicion is: such a low score is an indication that pipeline is broken and data preprocessing is not a solution. Could someone share any experience/opinion that might be helpful?</p>",
      "rawMarkdown": "Hello. \n\nI Preprocessed data - using <gap> where required; Removed !, ? signs where required; removed also [ ], ( ); I have been training model for 12 epochs; Test and val loss decreases for all epochs. But LB score is around 10.\n\nI could imagen: splitting examples in sentences would be helpful, have not done it yet.\n\nMy suspicion is: such a low score is an indication that pipeline is broken and data preprocessing is not a solution. Could someone share any experience/opinion that might be helpful?",
      "votes": null
    },
    {
      "id": "3417356",
      "postDate": "03/05/2026 07:42:38",
      "content": "<p>To narrow down my question: I have encountered several times opinion that this competition is mainly about preprocessing. So my question is: are RL based approaches necessary to achieve a score say around 30 or proper preprocessing is sufficient for this.</p>",
      "rawMarkdown": "To narrow down my question: I have encountered several times opinion that this competition is mainly about preprocessing. So my question is: are RL based approaches necessary to achieve a score say around 30 or proper preprocessing is sufficient for this.",
      "votes": null
    },
    {
      "id": "3417393",
      "postDate": "03/05/2026 09:15:57",
      "content": "<p>Look at the starter notebooks. They achieve close to 30 with no pre processing and very simple training. </p>",
      "rawMarkdown": "Look at the starter notebooks. They achieve close to 30 with no pre processing and very simple training.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3417356,
      "author_name": "davitkhantadze",
      "author_url": "",
      "post_date": "03/05/2026 07:42:38",
      "content": "<p>To narrow down my question: I have encountered several times opinion that this competition is mainly about preprocessing. So my question is: are RL based approaches necessary to achieve a score say around 30 or proper preprocessing is sufficient for this.</p>",
      "votes": null,
      "replies": [
        {
          "id": 3417393,
          "author_name": "tennogh",
          "author_url": "",
          "post_date": "03/05/2026 09:15:57",
          "content": "<p>Look at the starter notebooks. They achieve close to 30 with no pre processing and very simple training. </p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "3416608": "Hello. \n\nI Preprocessed data - using <gap> where required; Removed !, ? signs where required; removed also [ ], ( ); I have been training model for 12 epochs; Test and val loss decreases for all epochs. But LB score is around 10.\n\nI could imagen: splitting examples in sentences would be helpful, have not done it yet.\n\nMy suspicion is: such a low score is an indication that pipeline is broken and data preprocessing is not a solution. Could someone share any experience/opinion that might be helpful?",
    "3417356": "To narrow down my question: I have encountered several times opinion that this competition is mainly about preprocessing. So my question is: are RL based approaches necessary to achieve a score say around 30 or proper preprocessing is sufficient for this.",
    "3417393": "Look at the starter notebooks. They achieve close to 30 with no pre processing and very simple training."
  },
  "source": "meta"
}