{
  "id": 668812,
  "title": "Preprocessing",
  "url": "/competitions/deep-past-initiative-machine-translation/discussion/668812",
  "author_name": "",
  "post_date": "2026-01-19T06:35:30.028494900Z",
  "votes": null,
  "comment_count": 1,
  "views": 0,
  "content": "<p>I've got an idea on the post processing to match it up with the test dataset, but I'm still not sure on what to do pre processing wise to train a good model. Any like tips on where  to start? </p>",
  "messages": [
    {
      "id": "3393498",
      "postDate": "01/19/2026 06:35:30",
      "content": "<p>I've got an idea on the post processing to match it up with the test dataset, but I'm still not sure on what to do pre processing wise to train a good model. Any like tips on where  to start? </p>",
      "rawMarkdown": "I've got an idea on the post processing to match it up with the test dataset, but I'm still not sure on what to do pre processing wise to train a good model. Any like tips on where  to start?",
      "votes": null
    },
    {
      "id": "3395444",
      "postDate": "01/23/2026 01:59:27",
      "content": "<p>My preprocessing currently involves a custom data augmentation script that generates synthetic training examples by swapping proper nouns to prevent overfitting, followed by standard byte-level tokenization. Since I'm using ByT5, I'm keeping the rest of the text relatively raw/uncleaned to let the model handle the character-level nuances.</p>",
      "rawMarkdown": "My preprocessing currently involves a custom data augmentation script that generates synthetic training examples by swapping proper nouns to prevent overfitting, followed by standard byte-level tokenization. Since I'm using ByT5, I'm keeping the rest of the text relatively raw/uncleaned to let the model handle the character-level nuances.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3395444,
      "author_name": "leowithrow",
      "author_url": "",
      "post_date": "01/23/2026 01:59:27",
      "content": "<p>My preprocessing currently involves a custom data augmentation script that generates synthetic training examples by swapping proper nouns to prevent overfitting, followed by standard byte-level tokenization. Since I'm using ByT5, I'm keeping the rest of the text relatively raw/uncleaned to let the model handle the character-level nuances.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3393498": "I've got an idea on the post processing to match it up with the test dataset, but I'm still not sure on what to do pre processing wise to train a good model. Any like tips on where  to start?",
    "3395444": "My preprocessing currently involves a custom data augmentation script that generates synthetic training examples by swapping proper nouns to prevent overfitting, followed by standard byte-level tokenization. Since I'm using ByT5, I'm keeping the rest of the text relatively raw/uncleaned to let the model handle the character-level nuances."
  },
  "source": "meta"
}