{
  "id": 418316,
  "title": "Couldn't get the Best scores out for Matcha and Deplot",
  "url": "/competitions/benetech-making-graphs-accessible/discussion/418316",
  "author_name": "",
  "post_date": "2023-06-20T01:01:08.812094Z",
  "votes": 3,
  "comment_count": 1,
  "views": 0,
  "content": "<p>We tried to make use of both Matcha as well as DePlot but couldn't get them working to their full potential I feel (that's why waiting for solution notebooks to get public XD). Thanks <a href=\"https://www.kaggle.com/nbroad\" target=\"_blank\">@nbroad</a> for sharing his work in public. </p>\n<p>Our best scores were from deplot 0.58 public and 0.34 private (unfortunately we only selected the 0.33 ones). </p>\n<p>The strange thing about our work was our CV went from val_score = 0.62 to 0.71 but the public lb remain stuck at 0.58. Along with that  we tried to train the model on complete train data where the scores significantly deteriorated to 0.28 (pls anyone help me explain this). </p>\n<p>Along with that now I regret my work on scatter as I had made a separate train of another 50k images just of scatter but never submitted it 😔.</p>\n<p>Still, it was a great learning competition and keep on looking forward to more such comps. Also thanks <a href=\"https://www.kaggle.com/chaitanyagiri\" target=\"_blank\">@chaitanyagiri</a> for all his efforts in this comp.</p>",
  "messages": [
    {
      "id": "2309899",
      "postDate": "06/20/2023 01:01:08",
      "content": "<p>We tried to make use of both Matcha as well as DePlot but couldn't get them working to their full potential I feel (that's why waiting for solution notebooks to get public XD). Thanks <a href=\"https://www.kaggle.com/nbroad\" target=\"_blank\">@nbroad</a> for sharing his work in public. </p>\n<p>Our best scores were from deplot 0.58 public and 0.34 private (unfortunately we only selected the 0.33 ones). </p>\n<p>The strange thing about our work was our CV went from val_score = 0.62 to 0.71 but the public lb remain stuck at 0.58. Along with that  we tried to train the model on complete train data where the scores significantly deteriorated to 0.28 (pls anyone help me explain this). </p>\n<p>Along with that now I regret my work on scatter as I had made a separate train of another 50k images just of scatter but never submitted it 😔.</p>\n<p>Still, it was a great learning competition and keep on looking forward to more such comps. Also thanks <a href=\"https://www.kaggle.com/chaitanyagiri\" target=\"_blank\">@chaitanyagiri</a> for all his efforts in this comp.</p>",
      "rawMarkdown": "We tried to make use of both Matcha as well as DePlot but couldn't get them working to their full potential I feel (that's why waiting for solution notebooks to get public XD). Thanks @nbroad for sharing his work in public. \n\nOur best scores were from deplot 0.58 public and 0.34 private (unfortunately we only selected the 0.33 ones). \n\nThe strange thing about our work was our CV went from val_score = 0.62 to 0.71 but the public lb remain stuck at 0.58. Along with that  we tried to train the model on complete train data where the scores significantly deteriorated to 0.28 (pls anyone help me explain this). \n\nAlong with that now I regret my work on scatter as I had made a separate train of another 50k images just of scatter but never submitted it 😔.\n\nStill, it was a great learning competition and keep on looking forward to more such comps. Also thanks @chaitanyagiri for all his efforts in this comp.",
      "votes": null
    },
    {
      "id": "2310100",
      "postDate": "06/20/2023 05:38:44",
      "content": "<p><a href=\"https://www.kaggle.com/harshitkmr\" target=\"_blank\">@harshitkmr</a> thank you for feedback. I wait public notebooks from winners too.</p>",
      "rawMarkdown": "harshitkmr thank you for feedback. I wait public notebooks from winners too.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2310100,
      "author_name": "serangu",
      "author_url": "",
      "post_date": "06/20/2023 05:38:44",
      "content": "<p><a href=\"https://www.kaggle.com/harshitkmr\" target=\"_blank\">@harshitkmr</a> thank you for feedback. I wait public notebooks from winners too.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2309899": "We tried to make use of both Matcha as well as DePlot but couldn't get them working to their full potential I feel (that's why waiting for solution notebooks to get public XD). Thanks @nbroad for sharing his work in public. \n\nOur best scores were from deplot 0.58 public and 0.34 private (unfortunately we only selected the 0.33 ones). \n\nThe strange thing about our work was our CV went from val_score = 0.62 to 0.71 but the public lb remain stuck at 0.58. Along with that  we tried to train the model on complete train data where the scores significantly deteriorated to 0.28 (pls anyone help me explain this). \n\nAlong with that now I regret my work on scatter as I had made a separate train of another 50k images just of scatter but never submitted it 😔.\n\nStill, it was a great learning competition and keep on looking forward to more such comps. Also thanks @chaitanyagiri for all his efforts in this comp.",
    "2310100": "harshitkmr thank you for feedback. I wait public notebooks from winners too."
  },
  "source": "meta"
}