{
  "id": 106330,
  "title": "Impact of using QM9?",
  "url": "/competitions/champs-scalar-coupling/discussion/106330",
  "author_name": "",
  "post_date": "2019-08-29T06:31:24.306644Z",
  "votes": 2,
  "comment_count": 7,
  "views": 0,
  "content": "<p>Now that the competition is over, did anyone who used the QM9 dataset find that it gave them a boost or did it have a negligible effect?</p>\n\n<p>I didn’t use it but for my models because I found that my predictions of Mulliken charges from a NN were pretty accurate. Also <code>OpenBabel</code> has methods to calculate partial charges and dipole moments which is what I used.</p>\n\n<p>Interested to hear what other people thought of using the full QM9 dataset</p>",
  "messages": [
    {
      "id": "611170",
      "postDate": "08/29/2019 06:31:24",
      "content": "<p>Now that the competition is over, did anyone who used the QM9 dataset find that it gave them a boost or did it have a negligible effect?</p>\n\n<p>I didn’t use it but for my models because I found that my predictions of Mulliken charges from a NN were pretty accurate. Also <code>OpenBabel</code> has methods to calculate partial charges and dipole moments which is what I used.</p>\n\n<p>Interested to hear what other people thought of using the full QM9 dataset</p>",
      "rawMarkdown": "Now that the competition is over, did anyone who used the QM9 dataset find that it gave them a boost or did it have a negligible effect?\n\nI didn’t use it but for my models because I found that my predictions of Mulliken charges from a NN were pretty accurate. Also `OpenBabel` has methods to calculate partial charges and dipole moments which is what I used.\n\nInterested to hear what other people thought of using the full QM9 dataset",
      "votes": null
    },
    {
      "id": "611215",
      "postDate": "08/29/2019 07:02:54",
      "content": "<p>We've used predicted Mulliken charges up until second to last day of the competition, as we wanted to avoid using QM9. But things got very hot at the top, and we were not sure if others in top 5 have been using it, so we decided to switch to true Mullikens which did give us a significant boost, can't remember the exact number but we wouldn't be able to stay in top 5 without using it.</p>\n\n<p>Frankly, I think it was a bad decision to allow QM7/QM9 datasets. It's not in the spirit of the competition to use them, but they actually do help the scores, so in the end, these datasets force themselves on the competitors.</p>",
      "rawMarkdown": "We've used predicted Mulliken charges up until second to last day of the competition, as we wanted to avoid using QM9. But things got very hot at the top, and we were not sure if others in top 5 have been using it, so we decided to switch to true Mullikens which did give us a significant boost, can't remember the exact number but we wouldn't be able to stay in top 5 without using it.\n\nFrankly, I think it was a bad decision to allow QM7/QM9 datasets. It's not in the spirit of the competition to use them, but they actually do help the scores, so in the end, these datasets force themselves on the competitors.",
      "votes": null
    },
    {
      "id": "611236",
      "postDate": "08/29/2019 07:18:42",
      "content": "<p>Thanks for sharing. As a scientist I also didn’t want to use it in the spirit of the competition. Interesting to see how it helped</p>",
      "rawMarkdown": "Thanks for sharing. As a scientist I also didn’t want to use it in the spirit of the competition. Interesting to see how it helped",
      "votes": null
    },
    {
      "id": "611564",
      "postDate": "08/29/2019 10:20:51",
      "content": "<p>A teammate has dug out the numbers, we would have still been in top 5, but it did help. We were at -3.077 with predicted Mulliken, and hit -3.12 with QM9 Mulliken + another unrelated change that probably contributed a significant portion of the improvement.</p>",
      "rawMarkdown": "A teammate has dug out the numbers, we would have still been in top 5, but it did help. We were at -3.077 with predicted Mulliken, and hit -3.12 with QM9 Mulliken + another unrelated change that probably contributed a significant portion of the improvement.",
      "votes": null
    },
    {
      "id": "611742",
      "postDate": "08/29/2019 12:12:30",
      "content": "<p>We didn’t use QM7 or QM9, FWIW.  So while I don’t know how much it would have helped, it’s certainly still possible to do well without it.</p>",
      "rawMarkdown": "We didn’t use QM7 or QM9, FWIW.  So while I don’t know how much it would have helped, it’s certainly still possible to do well without it.",
      "votes": null
    },
    {
      "id": "612059",
      "postDate": "08/29/2019 15:27:28",
      "content": "<p>Using mulliken charges helps a bit.</p>",
      "rawMarkdown": "Using mulliken charges helps a bit.",
      "votes": null
    },
    {
      "id": "612102",
      "postDate": "08/29/2019 15:52:24",
      "content": "<p>This is consistent with the 0.05 uplift I saw when using mulliken the first time.</p>",
      "rawMarkdown": "This is consistent with the 0.05 uplift I saw when using mulliken the first time.",
      "votes": null
    },
    {
      "id": "612103",
      "postDate": "08/29/2019 15:53:08",
      "content": "<p>When will you say what you did use rather than what you didn't use?  ;)</p>",
      "rawMarkdown": "When will you say what you did use rather than what you didn't use?  ;)",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 611215,
      "author_name": "milosh",
      "author_url": "",
      "post_date": "08/29/2019 07:02:54",
      "content": "<p>We've used predicted Mulliken charges up until second to last day of the competition, as we wanted to avoid using QM9. But things got very hot at the top, and we were not sure if others in top 5 have been using it, so we decided to switch to true Mullikens which did give us a significant boost, can't remember the exact number but we wouldn't be able to stay in top 5 without using it.</p>\n\n<p>Frankly, I think it was a bad decision to allow QM7/QM9 datasets. It's not in the spirit of the competition to use them, but they actually do help the scores, so in the end, these datasets force themselves on the competitors.</p>",
      "votes": null,
      "replies": [
        {
          "id": 611236,
          "author_name": "anjum48",
          "author_url": "",
          "post_date": "08/29/2019 07:18:42",
          "content": "<p>Thanks for sharing. As a scientist I also didn’t want to use it in the spirit of the competition. Interesting to see how it helped</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 611564,
          "author_name": "milosh",
          "author_url": "",
          "post_date": "08/29/2019 10:20:51",
          "content": "<p>A teammate has dug out the numbers, we would have still been in top 5, but it did help. We were at -3.077 with predicted Mulliken, and hit -3.12 with QM9 Mulliken + another unrelated change that probably contributed a significant portion of the improvement.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 612102,
          "author_name": "cpmpml",
          "author_url": "",
          "post_date": "08/29/2019 15:52:24",
          "content": "<p>This is consistent with the 0.05 uplift I saw when using mulliken the first time.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 611742,
      "author_name": "zkolter",
      "author_url": "",
      "post_date": "08/29/2019 12:12:30",
      "content": "<p>We didn’t use QM7 or QM9, FWIW.  So while I don’t know how much it would have helped, it’s certainly still possible to do well without it.</p>",
      "votes": null,
      "replies": [
        {
          "id": 612103,
          "author_name": "cpmpml",
          "author_url": "",
          "post_date": "08/29/2019 15:53:08",
          "content": "<p>When will you say what you did use rather than what you didn't use?  ;)</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 612059,
      "author_name": "cpmpml",
      "author_url": "",
      "post_date": "08/29/2019 15:27:28",
      "content": "<p>Using mulliken charges helps a bit.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "611170": "Now that the competition is over, did anyone who used the QM9 dataset find that it gave them a boost or did it have a negligible effect?\n\nI didn’t use it but for my models because I found that my predictions of Mulliken charges from a NN were pretty accurate. Also `OpenBabel` has methods to calculate partial charges and dipole moments which is what I used.\n\nInterested to hear what other people thought of using the full QM9 dataset",
    "611215": "We've used predicted Mulliken charges up until second to last day of the competition, as we wanted to avoid using QM9. But things got very hot at the top, and we were not sure if others in top 5 have been using it, so we decided to switch to true Mullikens which did give us a significant boost, can't remember the exact number but we wouldn't be able to stay in top 5 without using it.\n\nFrankly, I think it was a bad decision to allow QM7/QM9 datasets. It's not in the spirit of the competition to use them, but they actually do help the scores, so in the end, these datasets force themselves on the competitors.",
    "611236": "Thanks for sharing. As a scientist I also didn’t want to use it in the spirit of the competition. Interesting to see how it helped",
    "611564": "A teammate has dug out the numbers, we would have still been in top 5, but it did help. We were at -3.077 with predicted Mulliken, and hit -3.12 with QM9 Mulliken + another unrelated change that probably contributed a significant portion of the improvement.",
    "611742": "We didn’t use QM7 or QM9, FWIW.  So while I don’t know how much it would have helped, it’s certainly still possible to do well without it.",
    "612059": "Using mulliken charges helps a bit.",
    "612102": "This is consistent with the 0.05 uplift I saw when using mulliken the first time.",
    "612103": "When will you say what you did use rather than what you didn't use?  ;)"
  },
  "source": "meta"
}