{
  "id": 287948,
  "title": "7th Place Solution",
  "content": "<p>Thanks kaggle and google research team for hosting this competition. Through this competition, I was able to gain a lot of experience and learned a lot. I am grateful for moved into gold zone. Feel free to leave comments.</p>\n<h3>TLDR</h3>\n<ol>\n<li>finetuning on squad2</li>\n<li>train use early stop</li>\n<li>hard voting ensemble</li>\n</ol>\n<h4>Models</h4>\n<p>I use 5 xlm-roberta-large, 5 muril, 4 rembert.</p>\n<h4>finetuning on squad2</h4>\n<p>deepset/xlm-roberta-large-squad2 was finetuning on squad2. In this point, i thought xlm-roberta-large-squad2 have good performance. Then finetune muril and rembert on squad2.</p>\n<h4>train</h4>\n<p>xlm-roberta-large-squad2: train 5 models with different seed (no fold out)</p>\n<p>muril: train 5 models with different fold</p>\n<p>rembert: train 4 models with different fold</p>\n<p>xlm-roberta-large train 2epochs with linear warm-up scheduler.</p>\n<p>muril and rembert train 10 epochs with linear warm-up scheduler and use ealry stop</p>\n<h4>hard voting ensemble</h4>\n<p>use hard voting ensemble across all models.</p>\n<p>xlm-roberta-large hard voting score: 0.757 / 0.728</p>\n<p>muril hard voting score: 0.764 / 0.754</p>\n<p>rembert hard voting score: 0.755 / 0.731</p>\n<p>all models hard voting score: 0.801 / 0.772</p>\n<p>Then I tuning hyper parameter and use other datasets (XQA_tamil, translated_tamil ,, etc)<br>\nI can improve score and my final submissions are as follows.</p>\n<p>all models hard voting score: 0.808 / 0.770</p>\n<p>all models + xlm-roberta-large(score base ensemble) + muril (score base ensemble): 0.808 / 0.770</p>\n<h4>did not work</h4>\n<ol>\n<li><p>score base ensemble</p></li>\n<li><p>muril, rembert seed ensemble</p></li>\n<li><p>google translate augmentation</p></li>\n<li><p>use more xlm-roberta-large models</p></li>\n</ol>\n<p>Congratulations on finishing the competition and see you next competitions.</p>",
  "comments": [
    {
      "id": 1590705,
      "author_name": "MINGOOK KIM",
      "author_url": "",
      "post_date": "2021-11-21T15:12:38.807000",
      "content": "<p>Congratulations on your achievements.</p>\n<p>Congratulations~</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 1585472,
      "author_name": "Azamat Tolegen",
      "author_url": "",
      "post_date": "2021-11-17T09:16:21.327000",
      "content": "<p>Congratulations! Planning to share ensembling sripts?</p>",
      "votes": 0,
      "replies": [
        {
          "id": 1585478,
          "author_name": "SeHwanJoo",
          "author_url": "",
          "post_date": "2021-11-17T09:23:14.473000",
          "content": "<p>Yes i will share soon.</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 1584828,
      "author_name": "Darek Kłeczek",
      "author_url": "",
      "post_date": "2021-11-16T19:41:55.250000",
      "content": "<p>Congratulations! Voting was a really effective technique here, I used it instead of blend in one of my final subs as well. </p>",
      "votes": 0,
      "replies": [
        {
          "id": 1585169,
          "author_name": "SeHwanJoo",
          "author_url": "",
          "post_date": "2021-11-17T06:15:40.143000",
          "content": "<p>Thanks. Congratulations for 1st place. Voting is powerful. And i saw your solution. Architecture and other technics are amazing!</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 1584213,
      "author_name": "Wonho Song",
      "author_url": "",
      "post_date": "2021-11-16T10:47:18.457000",
      "content": "<p>Congrats <a href=\"https://www.kaggle.com/sehwanjoo\" target=\"_blank\">@sehwanjoo</a> 🎉</p>\n<p>I have a question here: What is score based ensemble?</p>\n<p>Thanks for sharing your experiences.</p>",
      "votes": 0,
      "replies": [
        {
          "id": 1585162,
          "author_name": "SeHwanJoo",
          "author_url": "",
          "post_date": "2021-11-17T06:08:29.160000",
          "content": "<p>Thanks a lot. Xlm models have same tokenizer and add each model’s logits.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1584040,
      "author_name": "Luca Massaron",
      "author_url": "",
      "post_date": "2021-11-16T08:22:26.967000",
      "content": "<p>Congrats <a href=\"https://www.kaggle.com/sehwanjoo\" target=\"_blank\">@sehwanjoo</a>! I have a question, did you take the models at epoch completetion or at the best batch validation performance during the epoch?</p>",
      "votes": 0,
      "replies": [
        {
          "id": 1585161,
          "author_name": "SeHwanJoo",
          "author_url": "",
          "post_date": "2021-11-17T06:06:36.960000",
          "content": "<p>Thank you. I use early stop for muril and rembrrt. So i take the model at best <strong>epoch</strong> validation.</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 1585221,
          "author_name": "Luca Massaron",
          "author_url": "",
          "post_date": "2021-11-17T07:24:36.133000",
          "content": "<p>How did you set patience? I set it to 25% of an epoch but maybe I have been too restrained.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1585304,
          "author_name": "SeHwanJoo",
          "author_url": "",
          "post_date": "2021-11-17T08:23:44.340000",
          "content": "<p>I set patience 2epochs and take model from best jaccard validation epoch.</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 1583873,
      "author_name": "Raghawendra Singh",
      "author_url": "",
      "post_date": "2021-11-16T05:47:45.030000",
      "content": "<p>Congrats for gold!!!</p>",
      "votes": 0,
      "replies": [
        {
          "id": 1585157,
          "author_name": "SeHwanJoo",
          "author_url": "",
          "post_date": "2021-11-17T06:04:01.980000",
          "content": "<p>Thank you 👍👍👍</p>",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "comment_count": 13,
  "messages": [
    {
      "id": 1583782,
      "postDate": "2021-11-16T04:06:00.823Z",
      "content": "<p>Thanks kaggle and google research team for hosting this competition. Through this competition, I was able to gain a lot of experience and learned a lot. I am grateful for moved into gold zone. Feel free to leave comments.</p>\n<h3>TLDR</h3>\n<ol>\n<li>finetuning on squad2</li>\n<li>train use early stop</li>\n<li>hard voting ensemble</li>\n</ol>\n<h4>Models</h4>\n<p>I use 5 xlm-roberta-large, 5 muril, 4 rembert.</p>\n<h4>finetuning on squad2</h4>\n<p>deepset/xlm-roberta-large-squad2 was finetuning on squad2. In this point, i thought xlm-roberta-large-squad2 have good performance. Then finetune muril and rembert on squad2.</p>\n<h4>train</h4>\n<p>xlm-roberta-large-squad2: train 5 models with different seed (no fold out)</p>\n<p>muril: train 5 models with different fold</p>\n<p>rembert: train 4 models with different fold</p>\n<p>xlm-roberta-large train 2epochs with linear warm-up scheduler.</p>\n<p>muril and rembert train 10 epochs with linear warm-up scheduler and use ealry stop</p>\n<h4>hard voting ensemble</h4>\n<p>use hard voting ensemble across all models.</p>\n<p>xlm-roberta-large hard voting score: 0.757 / 0.728</p>\n<p>muril hard voting score: 0.764 / 0.754</p>\n<p>rembert hard voting score: 0.755 / 0.731</p>\n<p>all models hard voting score: 0.801 / 0.772</p>\n<p>Then I tuning hyper parameter and use other datasets (XQA_tamil, translated_tamil ,, etc)<br>\nI can improve score and my final submissions are as follows.</p>\n<p>all models hard voting score: 0.808 / 0.770</p>\n<p>all models + xlm-roberta-large(score base ensemble) + muril (score base ensemble): 0.808 / 0.770</p>\n<h4>did not work</h4>\n<ol>\n<li><p>score base ensemble</p></li>\n<li><p>muril, rembert seed ensemble</p></li>\n<li><p>google translate augmentation</p></li>\n<li><p>use more xlm-roberta-large models</p></li>\n</ol>\n<p>Congratulations on finishing the competition and see you next competitions.</p>",
      "rawMarkdown": "Thanks kaggle and google research team for hosting this competition. Through this competition, I was able to gain a lot of experience and learned a lot. I am grateful for moved into gold zone. Feel free to leave comments.\n\n### TLDR\n1. finetuning on squad2\n2. train use early stop\n3. hard voting ensemble\n\n#### Models\nI use 5 xlm-roberta-large, 5 muril, 4 rembert.\n\n#### finetuning on squad2\ndeepset/xlm-roberta-large-squad2 was finetuning on squad2. In this point, i thought xlm-roberta-large-squad2 have good performance. Then finetune muril and rembert on squad2.\n\n#### train\nxlm-roberta-large-squad2: train 5 models with different seed (no fold out)\n\nmuril: train 5 models with different fold\n\nrembert: train 4 models with different fold\n\nxlm-roberta-large train 2epochs with linear warm-up scheduler.\n\nmuril and rembert train 10 epochs with linear warm-up scheduler and use ealry stop\n\n#### hard voting ensemble\nuse hard voting ensemble across all models.\n\nxlm-roberta-large hard voting score: 0.757 / 0.728\n\nmuril hard voting score: 0.764 / 0.754\n\nrembert hard voting score: 0.755 / 0.731\n\nall models hard voting score: 0.801 / 0.772\n\n\nThen I tuning hyper parameter and use other datasets (XQA_tamil, translated_tamil ,, etc)\nI can improve score and my final submissions are as follows.\n\nall models hard voting score: 0.808 / 0.770\n\nall models + xlm-roberta-large(score base ensemble) + muril (score base ensemble): 0.808 / 0.770\n\n#### did not work\n1. score base ensemble\n\n2. muril, rembert seed ensemble\n\n3. google translate augmentation\n\n4. use more xlm-roberta-large models\n\nCongratulations on finishing the competition and see you next competitions.",
      "votes": 18
    },
    {
      "id": 1590705,
      "postDate": "2021-11-21T15:12:38.807Z",
      "content": "<p>Congratulations on your achievements.</p>\n<p>Congratulations~</p>",
      "rawMarkdown": "Congratulations on your achievements.\n\nCongratulations~",
      "votes": 1
    },
    {
      "id": 1585472,
      "postDate": "2021-11-17T09:16:21.327Z",
      "content": "<p>Congratulations! Planning to share ensembling sripts?</p>",
      "rawMarkdown": "Congratulations! Planning to share ensembling sripts?",
      "replies": [
        {
          "id": 1585478,
          "postDate": "2021-11-17T09:23:14.473Z",
          "content": "<p>Yes i will share soon.</p>",
          "rawMarkdown": "Yes i will share soon.",
          "votes": 1
        }
      ]
    },
    {
      "id": 1584828,
      "postDate": "2021-11-16T19:41:55.250Z",
      "content": "<p>Congratulations! Voting was a really effective technique here, I used it instead of blend in one of my final subs as well. </p>",
      "rawMarkdown": "Congratulations! Voting was a really effective technique here, I used it instead of blend in one of my final subs as well. ",
      "replies": [
        {
          "id": 1585169,
          "postDate": "2021-11-17T06:15:40.143Z",
          "content": "<p>Thanks. Congratulations for 1st place. Voting is powerful. And i saw your solution. Architecture and other technics are amazing!</p>",
          "rawMarkdown": "Thanks. Congratulations for 1st place. Voting is powerful. And i saw your solution. Architecture and other technics are amazing!",
          "votes": 1
        }
      ]
    },
    {
      "id": 1584213,
      "postDate": "2021-11-16T10:47:18.457Z",
      "content": "<p>Congrats <a href=\"https://www.kaggle.com/sehwanjoo\" target=\"_blank\">@sehwanjoo</a> 🎉</p>\n<p>I have a question here: What is score based ensemble?</p>\n<p>Thanks for sharing your experiences.</p>",
      "rawMarkdown": "Congrats @sehwanjoo 🎉\n\nI have a question here: What is score based ensemble?\n\nThanks for sharing your experiences.",
      "replies": [
        {
          "id": 1585162,
          "postDate": "2021-11-17T06:08:29.160Z",
          "content": "<p>Thanks a lot. Xlm models have same tokenizer and add each model’s logits.</p>",
          "rawMarkdown": "Thanks a lot. Xlm models have same tokenizer and add each model’s logits."
        }
      ]
    },
    {
      "id": 1584040,
      "postDate": "2021-11-16T08:22:26.967Z",
      "content": "<p>Congrats <a href=\"https://www.kaggle.com/sehwanjoo\" target=\"_blank\">@sehwanjoo</a>! I have a question, did you take the models at epoch completetion or at the best batch validation performance during the epoch?</p>",
      "rawMarkdown": "Congrats @sehwanjoo! I have a question, did you take the models at epoch completetion or at the best batch validation performance during the epoch?",
      "replies": [
        {
          "id": 1585161,
          "postDate": "2021-11-17T06:06:36.960Z",
          "content": "<p>Thank you. I use early stop for muril and rembrrt. So i take the model at best <strong>epoch</strong> validation.</p>",
          "rawMarkdown": "Thank you. I use early stop for muril and rembrrt. So i take the model at best **epoch** validation.",
          "votes": 1
        },
        {
          "id": 1585221,
          "postDate": "2021-11-17T07:24:36.133Z",
          "content": "<p>How did you set patience? I set it to 25% of an epoch but maybe I have been too restrained.</p>",
          "rawMarkdown": "How did you set patience? I set it to 25% of an epoch but maybe I have been too restrained."
        },
        {
          "id": 1585304,
          "postDate": "2021-11-17T08:23:44.340Z",
          "content": "<p>I set patience 2epochs and take model from best jaccard validation epoch.</p>",
          "rawMarkdown": "I set patience 2epochs and take model from best jaccard validation epoch.",
          "votes": 1
        }
      ]
    },
    {
      "id": 1583873,
      "postDate": "2021-11-16T05:47:45.030Z",
      "content": "<p>Congrats for gold!!!</p>",
      "rawMarkdown": "Congrats for gold!!!",
      "replies": [
        {
          "id": 1585157,
          "postDate": "2021-11-17T06:04:01.980Z",
          "content": "<p>Thank you 👍👍👍</p>",
          "rawMarkdown": "Thank you 👍👍👍"
        }
      ]
    }
  ],
  "complete": true,
  "source": "live",
  "fetched_at": "2026-10-19T16:21:48Z",
  "competition_id": 30060,
  "competition_slug": "chaii-hindi-and-tamil-question-answering",
  "team_id": 7412183,
  "solution_url": "/competitions/chaii-hindi-and-tamil-question-answering/writeups/sehwanjoo-7th-place-solution"
}