{
  "id": 352458,
  "title": "Good Ineffective - 6th top solution: simple and effective ",
  "content": "<p>Thanks for the Hosts organized this competation and all the teammates who equally shared the contribution to this competation.</p>\n<p>I'd like to share our solution, it is an easy and effective solution</p>\n<ol>\n<li>As like other top teams, the data preprcess is the key for this compettion. In the data,  we added element type before the element text, using the [SEP] token as a separator, then we integrate all the elements in an article into one big long sentence, and then predict the classification label for each sentence. the follwing is the sample of input data</li>\n</ol>\n<p>[SEP]Lead. <em>Discourse_00</em>[SEP]Position. <em>Discourse_01</em>[SEP]Claim. <em>Discourse_02</em> …….</p>\n<ol>\n<li><p>And then we used the deberta-base as the back-bone to test different model architecture, we used multi-drop out in the final output layer </p></li>\n<li><p>we chose the <strong>DeBERTa series model</strong>, specifically using <strong>\"microsoft/deberta-large \"</strong>, <strong>\"microsoft/deberta-v3-large \"</strong> from the HuggingFace library.</p></li>\n<li><p>AWP was also added in the training stage which was proven effective</p></li>\n<li><p>we also labeled the previous feedback1 training data as external data source which can also improved the cv &amp; pl</p></li>\n<li><p>Final CV-5Fold and simple weighted fusion.</p></li>\n</ol>\n<p>following are LB result </p>\n<ol>\n<li>deberta-base Baseline 5Fold，Public LB : 0.608；</li>\n<li>used the whole easay as the input ，Public LB : 0.589；</li>\n<li>used deberta-v3-large，Public LB : 0.577；</li>\n<li>introduced Pseudo Label，Public LB : 0.572；</li>\n<li>introduced AWP training，Public LB : 0.570；</li>\n<li>more fintune lr and awp arguments，Public LB : 0.568；</li>\n<li>simple average weight ensemble，(deberta-v3-large, deberta-large, deberta-xlarge)Public LB : 0.560；</li>\n</ol>",
  "comments": [
    {
      "id": 1949343,
      "author_name": "Gaurav Rawat",
      "author_url": "",
      "post_date": "2022-09-21T16:10:27.653000",
      "content": "<p>Congrats <a href=\"https://www.kaggle.com/leehann\" target=\"_blank\">@leehann</a>  simple and neat solution . you are always great at NLP comps 🙌</p>",
      "votes": 1,
      "replies": [
        {
          "id": 1949958,
          "author_name": "HZM",
          "author_url": "",
          "post_date": "2022-09-22T02:51:53.707000",
          "content": "<p>so much impressive by your great job on  amex </p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 1953774,
          "author_name": "Gaurav Rawat",
          "author_url": "",
          "post_date": "2022-09-24T18:08:02.377000",
          "content": "<p>Thanks <a href=\"https://www.kaggle.com/leehann\" target=\"_blank\">@leehann</a> 🙌😊</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 1954124,
          "author_name": "HZM",
          "author_url": "",
          "post_date": "2022-09-25T03:45:08.813000",
          "content": "<p>hope can work with you again in feedback 3</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1955075,
          "author_name": "Gaurav Rawat",
          "author_url": "",
          "post_date": "2022-09-25T17:39:55.263000",
          "content": "<p>Ya xiaoyi Is going.. Trying to do one comp at a time and taking a break from nlp ..currently concentrating on msci you are free to join us 😃. But yes can team up in future for feedback 4 or any new nlp comp 😄</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1955369,
          "author_name": "HZM",
          "author_url": "",
          "post_date": "2022-09-26T01:04:16.333000",
          "content": "<p>you are switiching to tabluar competions? hope learn with you in msci</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 1955386,
          "author_name": "Gaurav Rawat",
          "author_url": "",
          "post_date": "2022-09-26T01:25:00.880000",
          "content": "<p>Ha Ha lets see :) .. Not switching completely just wanted to learn and try something different so amex and msci .. next will come back post this .. </p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1945836,
      "author_name": "Psi",
      "author_url": "",
      "post_date": "2022-09-19T12:43:14.470000",
      "content": "<blockquote>\n  <p>we also labeled the previous feedback1 training data as external data source which can also improved the cv &amp; pl</p>\n</blockquote>\n<p>did you manually label that data?</p>",
      "votes": 0,
      "replies": [
        {
          "id": 1946626,
          "author_name": "HZM",
          "author_url": "",
          "post_date": "2022-09-20T01:06:13.217000",
          "content": "<p>firstly, we trained a deberta-large base on feedback2, and then use the trained model to label the feedback1 data.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1941612,
      "author_name": "SrihariMohan",
      "author_url": "",
      "post_date": "2022-09-16T07:36:31.613000",
      "content": "<p>Can you please share your solution?</p>",
      "votes": 0,
      "replies": []
    }
  ],
  "comment_count": 10,
  "messages": [
    {
      "id": 1938929,
      "postDate": "2022-09-14T12:45:06.193Z",
      "content": "<p>Thanks for the Hosts organized this competation and all the teammates who equally shared the contribution to this competation.</p>\n<p>I'd like to share our solution, it is an easy and effective solution</p>\n<ol>\n<li>As like other top teams, the data preprcess is the key for this compettion. In the data,  we added element type before the element text, using the [SEP] token as a separator, then we integrate all the elements in an article into one big long sentence, and then predict the classification label for each sentence. the follwing is the sample of input data</li>\n</ol>\n<p>[SEP]Lead. <em>Discourse_00</em>[SEP]Position. <em>Discourse_01</em>[SEP]Claim. <em>Discourse_02</em> …….</p>\n<ol>\n<li><p>And then we used the deberta-base as the back-bone to test different model architecture, we used multi-drop out in the final output layer </p></li>\n<li><p>we chose the <strong>DeBERTa series model</strong>, specifically using <strong>\"microsoft/deberta-large \"</strong>, <strong>\"microsoft/deberta-v3-large \"</strong> from the HuggingFace library.</p></li>\n<li><p>AWP was also added in the training stage which was proven effective</p></li>\n<li><p>we also labeled the previous feedback1 training data as external data source which can also improved the cv &amp; pl</p></li>\n<li><p>Final CV-5Fold and simple weighted fusion.</p></li>\n</ol>\n<p>following are LB result </p>\n<ol>\n<li>deberta-base Baseline 5Fold，Public LB : 0.608；</li>\n<li>used the whole easay as the input ，Public LB : 0.589；</li>\n<li>used deberta-v3-large，Public LB : 0.577；</li>\n<li>introduced Pseudo Label，Public LB : 0.572；</li>\n<li>introduced AWP training，Public LB : 0.570；</li>\n<li>more fintune lr and awp arguments，Public LB : 0.568；</li>\n<li>simple average weight ensemble，(deberta-v3-large, deberta-large, deberta-xlarge)Public LB : 0.560；</li>\n</ol>",
      "rawMarkdown": "Thanks for the Hosts organized this competation and all the teammates who equally shared the contribution to this competation.\n\nI'd like to share our solution, it is an easy and effective solution\n\n1. As like other top teams, the data preprcess is the key for this compettion. In the data,  we added element type before the element text, using the [SEP] token as a separator, then we integrate all the elements in an article into one big long sentence, and then predict the classification label for each sentence. the follwing is the sample of input data\n\n[SEP]Lead. *Discourse_00*[SEP]Position. *Discourse_01*[SEP]Claim. *Discourse_02* .......\n\n2. And then we used the deberta-base as the back-bone to test different model architecture, we used multi-drop out in the final output layer \n\n\n3. we chose the **DeBERTa series model**, specifically using **\"microsoft/deberta-large \"**, **\"microsoft/deberta-v3-large \"** from the HuggingFace library.\n\n4. AWP was also added in the training stage which was proven effective\n\n5. we also labeled the previous feedback1 training data as external data source which can also improved the cv & pl\n\n6. Final CV-5Fold and simple weighted fusion.\n\n\nfollowing are LB result \n1. deberta-base Baseline 5Fold，Public LB : 0.608；\n2. used the whole easay as the input ，Public LB : 0.589；\n3. used deberta-v3-large，Public LB : 0.577；\n4. introduced Pseudo Label，Public LB : 0.572；\n5. introduced AWP training，Public LB : 0.570；\n6. more fintune lr and awp arguments，Public LB : 0.568；\n7. simple average weight ensemble，(deberta-v3-large, deberta-large, deberta-xlarge)Public LB : 0.560；",
      "votes": 19
    },
    {
      "id": 1949343,
      "postDate": "2022-09-21T16:10:27.653Z",
      "content": "<p>Congrats <a href=\"https://www.kaggle.com/leehann\" target=\"_blank\">@leehann</a>  simple and neat solution . you are always great at NLP comps 🙌</p>",
      "rawMarkdown": "Congrats @leehann  simple and neat solution . you are always great at NLP comps 🙌",
      "votes": 1,
      "replies": [
        {
          "id": 1949958,
          "postDate": "2022-09-22T02:51:53.707Z",
          "content": "<p>so much impressive by your great job on  amex </p>",
          "rawMarkdown": "so much impressive by your great job on  amex ",
          "votes": 1
        },
        {
          "id": 1953774,
          "postDate": "2022-09-24T18:08:02.377Z",
          "content": "<p>Thanks <a href=\"https://www.kaggle.com/leehann\" target=\"_blank\">@leehann</a> 🙌😊</p>",
          "rawMarkdown": "Thanks @leehann 🙌😊",
          "votes": 1
        },
        {
          "id": 1954124,
          "postDate": "2022-09-25T03:45:08.813Z",
          "content": "<p>hope can work with you again in feedback 3</p>",
          "rawMarkdown": "hope can work with you again in feedback 3"
        },
        {
          "id": 1955075,
          "postDate": "2022-09-25T17:39:55.263Z",
          "content": "<p>Ya xiaoyi Is going.. Trying to do one comp at a time and taking a break from nlp ..currently concentrating on msci you are free to join us 😃. But yes can team up in future for feedback 4 or any new nlp comp 😄</p>",
          "rawMarkdown": "Ya xiaoyi Is going.. Trying to do one comp at a time and taking a break from nlp ..currently concentrating on msci you are free to join us 😃. But yes can team up in future for feedback 4 or any new nlp comp 😄"
        },
        {
          "id": 1955369,
          "postDate": "2022-09-26T01:04:16.333Z",
          "content": "<p>you are switiching to tabluar competions? hope learn with you in msci</p>",
          "rawMarkdown": "you are switiching to tabluar competions? hope learn with you in msci",
          "votes": 1
        },
        {
          "id": 1955386,
          "postDate": "2022-09-26T01:25:00.880Z",
          "content": "<p>Ha Ha lets see :) .. Not switching completely just wanted to learn and try something different so amex and msci .. next will come back post this .. </p>",
          "rawMarkdown": "Ha Ha lets see :) .. Not switching completely just wanted to learn and try something different so amex and msci .. next will come back post this .. "
        }
      ]
    },
    {
      "id": 1945836,
      "postDate": "2022-09-19T12:43:14.470Z",
      "content": "<blockquote>\n  <p>we also labeled the previous feedback1 training data as external data source which can also improved the cv &amp; pl</p>\n</blockquote>\n<p>did you manually label that data?</p>",
      "rawMarkdown": "> we also labeled the previous feedback1 training data as external data source which can also improved the cv & pl\n\ndid you manually label that data?",
      "replies": [
        {
          "id": 1946626,
          "postDate": "2022-09-20T01:06:13.217Z",
          "content": "<p>firstly, we trained a deberta-large base on feedback2, and then use the trained model to label the feedback1 data.</p>",
          "rawMarkdown": "firstly, we trained a deberta-large base on feedback2, and then use the trained model to label the feedback1 data."
        }
      ]
    },
    {
      "id": 1941612,
      "postDate": "2022-09-16T07:36:31.613Z",
      "content": "<p>Can you please share your solution?</p>",
      "rawMarkdown": "Can you please share your solution?"
    }
  ],
  "complete": true,
  "source": "live",
  "fetched_at": "2026-10-19T17:52:52Z",
  "competition_id": 35308,
  "competition_slug": "feedback-prize-effectiveness",
  "team_id": 8997842,
  "solution_url": "/competitions/feedback-prize-effectiveness/writeups/good-ineffective-good-ineffective-6th-top-solution"
}