{
  "id": 663505,
  "title": "Starter code and some tips",
  "url": "/competitions/deep-past-initiative-machine-translation/discussion/663505",
  "author_name": "",
  "post_date": "2025-12-18T12:01:52.908872200Z",
  "votes": 15,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Hi,</p>\n<p>I’m excited to participate in this very interesting competition.\nI’ve created a starter notebook and wanted to share it.</p>\n<p>Training: <a href=\"https://www.kaggle.com/code/takamichitoda/dpc-starter-train\" target=\"_blank\">https://www.kaggle.com/code/takamichitoda/dpc-starter-train</a></p>\n<p>Inference: <a href=\"https://www.kaggle.com/code/takamichitoda/dpc-starter-infer\" target=\"_blank\">https://www.kaggle.com/code/takamichitoda/dpc-starter-infer</a></p>\n<p>I’m currently experimenting with a simple sentence-alignment–style data augmentation\n(via <code>simple_sentence_aligner</code>), but so far the results are mixed.</p>\n<p>One interesting observation is that adding a task PREFIX to the input\nleads to a noticeable score improvement (+2.1).</p>",
  "messages": [
    {
      "id": "3378509",
      "postDate": "12/18/2025 12:01:52",
      "content": "<p>Hi,</p>\n<p>I’m excited to participate in this very interesting competition.\nI’ve created a starter notebook and wanted to share it.</p>\n<p>Training: <a href=\"https://www.kaggle.com/code/takamichitoda/dpc-starter-train\" target=\"_blank\">https://www.kaggle.com/code/takamichitoda/dpc-starter-train</a></p>\n<p>Inference: <a href=\"https://www.kaggle.com/code/takamichitoda/dpc-starter-infer\" target=\"_blank\">https://www.kaggle.com/code/takamichitoda/dpc-starter-infer</a></p>\n<p>I’m currently experimenting with a simple sentence-alignment–style data augmentation\n(via <code>simple_sentence_aligner</code>), but so far the results are mixed.</p>\n<p>One interesting observation is that adding a task PREFIX to the input\nleads to a noticeable score improvement (+2.1).</p>",
      "rawMarkdown": "Hi,\n\nI’m excited to participate in this very interesting competition.\nI’ve created a starter notebook and wanted to share it.\n\nTraining: https://www.kaggle.com/code/takamichitoda/dpc-starter-train\n\nInference: https://www.kaggle.com/code/takamichitoda/dpc-starter-infer\n\n\nI’m currently experimenting with a simple sentence-alignment–style data augmentation\n(via `simple_sentence_aligner`), but so far the results are mixed.\n\nOne interesting observation is that adding a task PREFIX to the input\nleads to a noticeable score improvement (+2.1).",
      "votes": null
    },
    {
      "id": "3378523",
      "postDate": "12/18/2025 12:27:05",
      "content": "<p>I guess what you want to explain with the prefix method is something like the system prompt in LLMs, right?</p>",
      "rawMarkdown": "I guess what you want to explain with the prefix method is something like the system prompt in LLMs, right?",
      "votes": null
    },
    {
      "id": "3378542",
      "postDate": "12/18/2025 13:01:33",
      "content": "<p>Yes, it’s similar in that it sets the context for the model.\nIn T5-style models, the prefix is just normal text seen during training, not a special system-level mechanism.</p>\n<p>It helps the model recognize that the input should be treated as a translation task.</p>",
      "rawMarkdown": "Yes, it’s similar in that it sets the context for the model.\nIn T5-style models, the prefix is just normal text seen during training, not a special system-level mechanism.\n\nIt helps the model recognize that the input should be treated as a translation task.",
      "votes": null
    },
    {
      "id": "3378630",
      "postDate": "12/18/2025 15:37:53",
      "content": "<p>Another point is that the current maximum length of <strong>translation</strong> is 512, and some <strong>transliteration</strong> are not aligned. Therefore, increasing the length can reduce the value of  <strong>BP=exp(1−r/c)</strong> in the evaluation metric <strong>BLEU</strong>, thereby decreasing the penalty strength and improving the score.Good luck!:)</p>",
      "rawMarkdown": "Another point is that the current maximum length of **translation** is 512, and some **transliteration** are not aligned. Therefore, increasing the length can reduce the value of  **BP=exp(1−r/c)** in the evaluation metric **BLEU**, thereby decreasing the penalty strength and improving the score.Good luck!:)",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3378523,
      "author_name": "ulasdesouza",
      "author_url": "",
      "post_date": "12/18/2025 12:27:05",
      "content": "<p>I guess what you want to explain with the prefix method is something like the system prompt in LLMs, right?</p>",
      "votes": null,
      "replies": [
        {
          "id": 3378542,
          "author_name": "takamichitoda",
          "author_url": "",
          "post_date": "12/18/2025 13:01:33",
          "content": "<p>Yes, it’s similar in that it sets the context for the model.\nIn T5-style models, the prefix is just normal text seen during training, not a special system-level mechanism.</p>\n<p>It helps the model recognize that the input should be treated as a translation task.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 3378630,
      "author_name": "qifeihhh666",
      "author_url": "",
      "post_date": "12/18/2025 15:37:53",
      "content": "<p>Another point is that the current maximum length of <strong>translation</strong> is 512, and some <strong>transliteration</strong> are not aligned. Therefore, increasing the length can reduce the value of  <strong>BP=exp(1−r/c)</strong> in the evaluation metric <strong>BLEU</strong>, thereby decreasing the penalty strength and improving the score.Good luck!:)</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3378509": "Hi,\n\nI’m excited to participate in this very interesting competition.\nI’ve created a starter notebook and wanted to share it.\n\nTraining: https://www.kaggle.com/code/takamichitoda/dpc-starter-train\n\nInference: https://www.kaggle.com/code/takamichitoda/dpc-starter-infer\n\n\nI’m currently experimenting with a simple sentence-alignment–style data augmentation\n(via `simple_sentence_aligner`), but so far the results are mixed.\n\nOne interesting observation is that adding a task PREFIX to the input\nleads to a noticeable score improvement (+2.1).",
    "3378523": "I guess what you want to explain with the prefix method is something like the system prompt in LLMs, right?",
    "3378542": "Yes, it’s similar in that it sets the context for the model.\nIn T5-style models, the prefix is just normal text seen during training, not a special system-level mechanism.\n\nIt helps the model recognize that the input should be treated as a translation task.",
    "3378630": "Another point is that the current maximum length of **translation** is 512, and some **transliteration** are not aligned. Therefore, increasing the length can reduce the value of  **BP=exp(1−r/c)** in the evaluation metric **BLEU**, thereby decreasing the penalty strength and improving the score.Good luck!:)"
  },
  "source": "meta"
}