{
  "id": 230528,
  "title": "Public/Private split",
  "url": "/competitions/bms-molecular-translation/discussion/230528",
  "author_name": "",
  "post_date": "2021-04-04T08:42:49.031751700Z",
  "votes": 6,
  "comment_count": 7,
  "views": 0,
  "content": "<p>Was the public and private test sets split randomly from the union of them?</p>\n<p>I can only see 7 molecules in the training data that has no carbon in them, namely:</p>\n<pre><code>           image_id        InChI\n82797      08bc4cbc516a    InChI=1S/Cl11FN6P6/c1-19(2)13-20(3,4)15-22(7,8...\n1308353    8a09da6e62d1    InChI=1S/BCl7H2N2Si2/c2-1(9-11(3,4)5)10-12(6,7...\n1335903    8cecf679c628    InChI=1S/Cl5HN3OP3/c1-10(2)6-11(3,4)8-12(5,9)7...\n1463871    9a838b52519b    InChI=1S/Cl4FN3OP2S/c1-10(2)6-11(3,4)8-12(5,9)...\n1475838    9bcb74107a90    InChI=1S/Cl8H2N5O2P5/c1-16(2)9-17(3,4)11-20(10...\n1507188    9f1b6ebfc567    InChI=1S/Cl13N5OP6/c1-20(2,3)14-21(4,5)15-22(6...\n1676354    b0fc45c025b8    InChI=1S/Cl5FN3P3/c1-10(2)7-11(3,4)9-12(5,6)8-10\n</code></pre>\n<p>Are we supposed to take additional care of cases where there is no carbon present, or is the private set drawn from the same distribution as the public one? (So that, is it reasonable to expect the private score somewhat close to the public score because of same distribution?)</p>",
  "messages": [
    {
      "id": "1262420",
      "postDate": "04/04/2021 08:42:49",
      "content": "<p>Was the public and private test sets split randomly from the union of them?</p>\n<p>I can only see 7 molecules in the training data that has no carbon in them, namely:</p>\n<pre><code>           image_id        InChI\n82797      08bc4cbc516a    InChI=1S/Cl11FN6P6/c1-19(2)13-20(3,4)15-22(7,8...\n1308353    8a09da6e62d1    InChI=1S/BCl7H2N2Si2/c2-1(9-11(3,4)5)10-12(6,7...\n1335903    8cecf679c628    InChI=1S/Cl5HN3OP3/c1-10(2)6-11(3,4)8-12(5,9)7...\n1463871    9a838b52519b    InChI=1S/Cl4FN3OP2S/c1-10(2)6-11(3,4)8-12(5,9)...\n1475838    9bcb74107a90    InChI=1S/Cl8H2N5O2P5/c1-16(2)9-17(3,4)11-20(10...\n1507188    9f1b6ebfc567    InChI=1S/Cl13N5OP6/c1-20(2,3)14-21(4,5)15-22(6...\n1676354    b0fc45c025b8    InChI=1S/Cl5FN3P3/c1-10(2)7-11(3,4)9-12(5,6)8-10\n</code></pre>\n<p>Are we supposed to take additional care of cases where there is no carbon present, or is the private set drawn from the same distribution as the public one? (So that, is it reasonable to expect the private score somewhat close to the public score because of same distribution?)</p>",
      "rawMarkdown": "Was the public and private test sets split randomly from the union of them?\n\nI can only see 7 molecules in the training data that has no carbon in them, namely:\n```\n\t       image_id\t       InChI\n82797\t  08bc4cbc516a\t  InChI=1S/Cl11FN6P6/c1-19(2)13-20(3,4)15-22(7,8...\n1308353\t8a09da6e62d1\tInChI=1S/BCl7H2N2Si2/c2-1(9-11(3,4)5)10-12(6,7...\n1335903\t8cecf679c628\tInChI=1S/Cl5HN3OP3/c1-10(2)6-11(3,4)8-12(5,9)7...\n1463871\t9a838b52519b\tInChI=1S/Cl4FN3OP2S/c1-10(2)6-11(3,4)8-12(5,9)...\n1475838\t9bcb74107a90\tInChI=1S/Cl8H2N5O2P5/c1-16(2)9-17(3,4)11-20(10...\n1507188\t9f1b6ebfc567\tInChI=1S/Cl13N5OP6/c1-20(2,3)14-21(4,5)15-22(6...\n1676354\tb0fc45c025b8\tInChI=1S/Cl5FN3P3/c1-10(2)7-11(3,4)9-12(5,6)8-10\n```\n\nAre we supposed to take additional care of cases where there is no carbon present, or is the private set drawn from the same distribution as the public one? (So that, is it reasonable to expect the private score somewhat close to the public score because of same distribution?)",
      "votes": null
    },
    {
      "id": "1267659",
      "postDate": "04/08/2021 17:19:20",
      "content": "<p>I would like to know in order to make inference and submit only on public portion.</p>",
      "rawMarkdown": "I would like to know in order to make inference and submit only on public portion.",
      "votes": null
    },
    {
      "id": "1267682",
      "postDate": "04/08/2021 17:43:11",
      "content": "<p>That didn't even cross my mind. That would be cool if we could do that!</p>",
      "rawMarkdown": "That didn't even cross my mind. That would be cool if we could do that!",
      "votes": null
    },
    {
      "id": "1267704",
      "postDate": "04/08/2021 17:58:41",
      "content": "<p><a href=\"https://www.kaggle.com/inversion\" target=\"_blank\">@inversion</a> <strong>Can we ignore non-organic molecules without being disqualified</strong> because of not being general enough?<br>\nTo train an autoregressive model on this little non-organic data probably hurts. But as the test's distribution seems to be much the same as of the training data based on LB scores, I would like to just ignore non-organic molecules. But I fear that doing so might bring us disqualification.</p>\n<p>In the external \"InChI only data\" there also are only 29 labels that are non-organic. (10M/2.5M-&gt;4; 29/7-&gt;~4 seems like the additional data is of the same distribution, too, in this regard at least)</p>\n<p>Developing something alongside a simple autoregressive task is too much work just to be general. Instead, if we are supposed to be robust not only on organic molecules, it's so much easier to just add more non-organic ones to the training data.</p>\n<p>These are all the molecules with no carbon in them in the external data.</p>\n<pre><code>Name       InChI\n763817     InChI=1S/Cl12N6P6/c1-19(2,3)13-23(11)15-21(7,8...\n963216     InChI=1S/Cl16N8P8/c1-25(2,3)17-31(18-26(4,5)6)...\n1188210    InChI=1S/Cl2F2N3O2PS2/c1-10(2)5-11(3,8)7-12(4,...\n1907261    InChI=1S/Cl2O5S2/c1-8(3,4)7-9(2,5)6\n2029843    InChI=1S/Cl3F3OSi2/c1-8(2,3)7-9(4,5)6\n2351320    InChI=1S/Br2Cl4N3P3/c1-10(3)7-11(2,4)9-12(5,6)...\n2583289    InChI=1S/Cl8N4P4/c1-13(2,3)9-16(8)11-14(4,5)10...\n2764125    InChI=1S/Cl5F3N4P4/c1-13(2)9-14(3,4)11-16(7,8)...\n3021362    InChI=1S/Cl6F3NSi3/c1-11(2,3)10(12(4,5)7)13(6,8)9\n3342899    InChI=1S/BrClF4N3P3/c1-10(3)7-11(2,4)9-12(5,6)...\n3654353    InChI=1S/Cl4NO5PS2/c1-11(2,5-12(3,6)7)10-13(4,8)9\n4466221    InChI=1S/Cl7N2P3S/c1-10(2,3)8-11(4,5)9-12(6,7)13\n4612505    InChI=1S/Cl10N4O2P4S/c1-17(2,3)11-19(7,8)13-21...\n5471325    InChI=1S/Cl2O4Si4/c1-10(2)5-8-3-7-4-9-6-10\n6141923    InChI=1S/Cl3IO12/c5-1(6,7)14-4(15-2(8,9)10)16-...\n6443167    InChI=1S/Cl7NSi2/c1-8(9(2,3)4)10(5,6)7\n6731251    InChI=1S/Cl11N4OP5/c1-17(2,3)12-18(4,5)13-19(6...\n7242878    InChI=1S/ClFHNO4S2/c1-8(4,5)3-9(2,6)7/h3H\n7310492    InChI=1S/ClFO5S2/c1-8(3,4)7-9(2,5)6\n8129600    InChI=1S/Cl12O8Si7/c1-21(2)13-22(3,4)16-26(11)...\n8149394    InChI=1S/Cl4NO2PS/c1-8(2,3)5-9(4,6)7\n9060183    InChI=1S/Cl6NSi2/c1-7(8(2)3)9(4,5)6\n9274099    InChI=1S/Cl3H2N4O2PS2/c1-10(4)5-11(2,8)7-12(3,...\n9377683    InChI=1S/ClF3NO2PS/c1-9(6,7)5-8(2,3)4\n9447617    InChI=1S/Cl6N3P3/c1-8-10(3)7-12(5,6)9(2)11(8)4\n9570835    InChI=1S/Cl3N3O3S3/c1-10(7)4-11(2,8)6-12(3,9)5-10\n9582304    InChI=1S/Br4Cl2N3P3/c1-10(2)7-11(3,5)9-12(4,6)...\n9852973    InChI=1S/ClHO4/c2-1(3,4)5/h(H,2,3,4,5)/i2+2,3+...\n</code></pre>",
      "rawMarkdown": "inversion **Can we ignore non-organic molecules without being disqualified** because of not being general enough?\nTo train an autoregressive model on this little non-organic data probably hurts. But as the test's distribution seems to be much the same as of the training data based on LB scores, I would like to just ignore non-organic molecules. But I fear that doing so might bring us disqualification.\n\nIn the external \"InChI only data\" there also are only 29 labels that are non-organic. (10M/2.5M->4; 29/7->~4 seems like the additional data is of the same distribution, too, in this regard at least)\n\nDeveloping something alongside a simple autoregressive task is too much work just to be general. Instead, if we are supposed to be robust not only on organic molecules, it's so much easier to just add more non-organic ones to the training data.\n\nThese are all the molecules with no carbon in them in the external data.\n\n```\nName       InChI\n763817\t InChI=1S/Cl12N6P6/c1-19(2,3)13-23(11)15-21(7,8...\n963216\t InChI=1S/Cl16N8P8/c1-25(2,3)17-31(18-26(4,5)6)...\n1188210\tInChI=1S/Cl2F2N3O2PS2/c1-10(2)5-11(3,8)7-12(4,...\n1907261\tInChI=1S/Cl2O5S2/c1-8(3,4)7-9(2,5)6\n2029843\tInChI=1S/Cl3F3OSi2/c1-8(2,3)7-9(4,5)6\n2351320\tInChI=1S/Br2Cl4N3P3/c1-10(3)7-11(2,4)9-12(5,6)...\n2583289\tInChI=1S/Cl8N4P4/c1-13(2,3)9-16(8)11-14(4,5)10...\n2764125\tInChI=1S/Cl5F3N4P4/c1-13(2)9-14(3,4)11-16(7,8)...\n3021362\tInChI=1S/Cl6F3NSi3/c1-11(2,3)10(12(4,5)7)13(6,8)9\n3342899\tInChI=1S/BrClF4N3P3/c1-10(3)7-11(2,4)9-12(5,6)...\n3654353\tInChI=1S/Cl4NO5PS2/c1-11(2,5-12(3,6)7)10-13(4,8)9\n4466221\tInChI=1S/Cl7N2P3S/c1-10(2,3)8-11(4,5)9-12(6,7)13\n4612505\tInChI=1S/Cl10N4O2P4S/c1-17(2,3)11-19(7,8)13-21...\n5471325\tInChI=1S/Cl2O4Si4/c1-10(2)5-8-3-7-4-9-6-10\n6141923\tInChI=1S/Cl3IO12/c5-1(6,7)14-4(15-2(8,9)10)16-...\n6443167\tInChI=1S/Cl7NSi2/c1-8(9(2,3)4)10(5,6)7\n6731251\tInChI=1S/Cl11N4OP5/c1-17(2,3)12-18(4,5)13-19(6...\n7242878\tInChI=1S/ClFHNO4S2/c1-8(4,5)3-9(2,6)7/h3H\n7310492\tInChI=1S/ClFO5S2/c1-8(3,4)7-9(2,5)6\n8129600\tInChI=1S/Cl12O8Si7/c1-21(2)13-22(3,4)16-26(11)...\n8149394\tInChI=1S/Cl4NO2PS/c1-8(2,3)5-9(4,6)7\n9060183\tInChI=1S/Cl6NSi2/c1-7(8(2)3)9(4,5)6\n9274099\tInChI=1S/Cl3H2N4O2PS2/c1-10(4)5-11(2,8)7-12(3,...\n9377683\tInChI=1S/ClF3NO2PS/c1-9(6,7)5-8(2,3)4\n9447617\tInChI=1S/Cl6N3P3/c1-8-10(3)7-12(5,6)9(2)11(8)4\n9570835\tInChI=1S/Cl3N3O3S3/c1-10(7)4-11(2,8)6-12(3,9)5-10\n9582304\tInChI=1S/Br4Cl2N3P3/c1-10(2)7-11(3,5)9-12(4,6)...\n9852973\tInChI=1S/ClHO4/c2-1(3,4)5/h(H,2,3,4,5)/i2+2,3+...\n```",
      "votes": null
    },
    {
      "id": "1267723",
      "postDate": "04/08/2021 18:12:14",
      "content": "<p>You are absolutely free to ignore as much as the training data as you see fit!</p>",
      "rawMarkdown": "You are absolutely free to ignore as much as the training data as you see fit!",
      "votes": null
    },
    {
      "id": "1267726",
      "postDate": "04/08/2021 18:13:24",
      "content": "<p>Thank you!</p>",
      "rawMarkdown": "Thank you!",
      "votes": null
    },
    {
      "id": "1267735",
      "postDate": "04/08/2021 18:21:09",
      "content": "<p>My concern was that a model that is only able to predict organic molecules will be discarded regardless of the LB score because it must be able to work on those too in the real world where it will be used.</p>\n<p>But <strong>as I understand</strong> now, this is not required and <strong>only LB matters</strong> (as long as all rules are met).</p>",
      "rawMarkdown": "My concern was that a model that is only able to predict organic molecules will be discarded regardless of the LB score because it must be able to work on those too in the real world where it will be used.\n\nBut **as I understand** now, this is not required and **only LB matters** (as long as all rules are met).",
      "votes": null
    },
    {
      "id": "1267797",
      "postDate": "04/08/2021 20:03:48",
      "content": "<p>Till now, I wrongly called \"carbon-free\" molecules as equal to \"non-organic\" ones.</p>",
      "rawMarkdown": "Till now, I wrongly called \"carbon-free\" molecules as equal to \"non-organic\" ones.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1267659,
      "author_name": "claverru",
      "author_url": "",
      "post_date": "04/08/2021 17:19:20",
      "content": "<p>I would like to know in order to make inference and submit only on public portion.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1267682,
          "author_name": "nofreewill",
          "author_url": "",
          "post_date": "04/08/2021 17:43:11",
          "content": "<p>That didn't even cross my mind. That would be cool if we could do that!</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1267704,
      "author_name": "nofreewill",
      "author_url": "",
      "post_date": "04/08/2021 17:58:41",
      "content": "<p><a href=\"https://www.kaggle.com/inversion\" target=\"_blank\">@inversion</a> <strong>Can we ignore non-organic molecules without being disqualified</strong> because of not being general enough?<br>\nTo train an autoregressive model on this little non-organic data probably hurts. But as the test's distribution seems to be much the same as of the training data based on LB scores, I would like to just ignore non-organic molecules. But I fear that doing so might bring us disqualification.</p>\n<p>In the external \"InChI only data\" there also are only 29 labels that are non-organic. (10M/2.5M-&gt;4; 29/7-&gt;~4 seems like the additional data is of the same distribution, too, in this regard at least)</p>\n<p>Developing something alongside a simple autoregressive task is too much work just to be general. Instead, if we are supposed to be robust not only on organic molecules, it's so much easier to just add more non-organic ones to the training data.</p>\n<p>These are all the molecules with no carbon in them in the external data.</p>\n<pre><code>Name       InChI\n763817     InChI=1S/Cl12N6P6/c1-19(2,3)13-23(11)15-21(7,8...\n963216     InChI=1S/Cl16N8P8/c1-25(2,3)17-31(18-26(4,5)6)...\n1188210    InChI=1S/Cl2F2N3O2PS2/c1-10(2)5-11(3,8)7-12(4,...\n1907261    InChI=1S/Cl2O5S2/c1-8(3,4)7-9(2,5)6\n2029843    InChI=1S/Cl3F3OSi2/c1-8(2,3)7-9(4,5)6\n2351320    InChI=1S/Br2Cl4N3P3/c1-10(3)7-11(2,4)9-12(5,6)...\n2583289    InChI=1S/Cl8N4P4/c1-13(2,3)9-16(8)11-14(4,5)10...\n2764125    InChI=1S/Cl5F3N4P4/c1-13(2)9-14(3,4)11-16(7,8)...\n3021362    InChI=1S/Cl6F3NSi3/c1-11(2,3)10(12(4,5)7)13(6,8)9\n3342899    InChI=1S/BrClF4N3P3/c1-10(3)7-11(2,4)9-12(5,6)...\n3654353    InChI=1S/Cl4NO5PS2/c1-11(2,5-12(3,6)7)10-13(4,8)9\n4466221    InChI=1S/Cl7N2P3S/c1-10(2,3)8-11(4,5)9-12(6,7)13\n4612505    InChI=1S/Cl10N4O2P4S/c1-17(2,3)11-19(7,8)13-21...\n5471325    InChI=1S/Cl2O4Si4/c1-10(2)5-8-3-7-4-9-6-10\n6141923    InChI=1S/Cl3IO12/c5-1(6,7)14-4(15-2(8,9)10)16-...\n6443167    InChI=1S/Cl7NSi2/c1-8(9(2,3)4)10(5,6)7\n6731251    InChI=1S/Cl11N4OP5/c1-17(2,3)12-18(4,5)13-19(6...\n7242878    InChI=1S/ClFHNO4S2/c1-8(4,5)3-9(2,6)7/h3H\n7310492    InChI=1S/ClFO5S2/c1-8(3,4)7-9(2,5)6\n8129600    InChI=1S/Cl12O8Si7/c1-21(2)13-22(3,4)16-26(11)...\n8149394    InChI=1S/Cl4NO2PS/c1-8(2,3)5-9(4,6)7\n9060183    InChI=1S/Cl6NSi2/c1-7(8(2)3)9(4,5)6\n9274099    InChI=1S/Cl3H2N4O2PS2/c1-10(4)5-11(2,8)7-12(3,...\n9377683    InChI=1S/ClF3NO2PS/c1-9(6,7)5-8(2,3)4\n9447617    InChI=1S/Cl6N3P3/c1-8-10(3)7-12(5,6)9(2)11(8)4\n9570835    InChI=1S/Cl3N3O3S3/c1-10(7)4-11(2,8)6-12(3,9)5-10\n9582304    InChI=1S/Br4Cl2N3P3/c1-10(2)7-11(3,5)9-12(4,6)...\n9852973    InChI=1S/ClHO4/c2-1(3,4)5/h(H,2,3,4,5)/i2+2,3+...\n</code></pre>",
      "votes": null,
      "replies": [
        {
          "id": 1267723,
          "author_name": "inversion",
          "author_url": "",
          "post_date": "04/08/2021 18:12:14",
          "content": "<p>You are absolutely free to ignore as much as the training data as you see fit!</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1267726,
          "author_name": "nofreewill",
          "author_url": "",
          "post_date": "04/08/2021 18:13:24",
          "content": "<p>Thank you!</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1267735,
          "author_name": "nofreewill",
          "author_url": "",
          "post_date": "04/08/2021 18:21:09",
          "content": "<p>My concern was that a model that is only able to predict organic molecules will be discarded regardless of the LB score because it must be able to work on those too in the real world where it will be used.</p>\n<p>But <strong>as I understand</strong> now, this is not required and <strong>only LB matters</strong> (as long as all rules are met).</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1267797,
      "author_name": "nofreewill",
      "author_url": "",
      "post_date": "04/08/2021 20:03:48",
      "content": "<p>Till now, I wrongly called \"carbon-free\" molecules as equal to \"non-organic\" ones.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1262420": "Was the public and private test sets split randomly from the union of them?\n\nI can only see 7 molecules in the training data that has no carbon in them, namely:\n```\n\t       image_id\t       InChI\n82797\t  08bc4cbc516a\t  InChI=1S/Cl11FN6P6/c1-19(2)13-20(3,4)15-22(7,8...\n1308353\t8a09da6e62d1\tInChI=1S/BCl7H2N2Si2/c2-1(9-11(3,4)5)10-12(6,7...\n1335903\t8cecf679c628\tInChI=1S/Cl5HN3OP3/c1-10(2)6-11(3,4)8-12(5,9)7...\n1463871\t9a838b52519b\tInChI=1S/Cl4FN3OP2S/c1-10(2)6-11(3,4)8-12(5,9)...\n1475838\t9bcb74107a90\tInChI=1S/Cl8H2N5O2P5/c1-16(2)9-17(3,4)11-20(10...\n1507188\t9f1b6ebfc567\tInChI=1S/Cl13N5OP6/c1-20(2,3)14-21(4,5)15-22(6...\n1676354\tb0fc45c025b8\tInChI=1S/Cl5FN3P3/c1-10(2)7-11(3,4)9-12(5,6)8-10\n```\n\nAre we supposed to take additional care of cases where there is no carbon present, or is the private set drawn from the same distribution as the public one? (So that, is it reasonable to expect the private score somewhat close to the public score because of same distribution?)",
    "1267659": "I would like to know in order to make inference and submit only on public portion.",
    "1267682": "That didn't even cross my mind. That would be cool if we could do that!",
    "1267704": "inversion **Can we ignore non-organic molecules without being disqualified** because of not being general enough?\nTo train an autoregressive model on this little non-organic data probably hurts. But as the test's distribution seems to be much the same as of the training data based on LB scores, I would like to just ignore non-organic molecules. But I fear that doing so might bring us disqualification.\n\nIn the external \"InChI only data\" there also are only 29 labels that are non-organic. (10M/2.5M->4; 29/7->~4 seems like the additional data is of the same distribution, too, in this regard at least)\n\nDeveloping something alongside a simple autoregressive task is too much work just to be general. Instead, if we are supposed to be robust not only on organic molecules, it's so much easier to just add more non-organic ones to the training data.\n\nThese are all the molecules with no carbon in them in the external data.\n\n```\nName       InChI\n763817\t InChI=1S/Cl12N6P6/c1-19(2,3)13-23(11)15-21(7,8...\n963216\t InChI=1S/Cl16N8P8/c1-25(2,3)17-31(18-26(4,5)6)...\n1188210\tInChI=1S/Cl2F2N3O2PS2/c1-10(2)5-11(3,8)7-12(4,...\n1907261\tInChI=1S/Cl2O5S2/c1-8(3,4)7-9(2,5)6\n2029843\tInChI=1S/Cl3F3OSi2/c1-8(2,3)7-9(4,5)6\n2351320\tInChI=1S/Br2Cl4N3P3/c1-10(3)7-11(2,4)9-12(5,6)...\n2583289\tInChI=1S/Cl8N4P4/c1-13(2,3)9-16(8)11-14(4,5)10...\n2764125\tInChI=1S/Cl5F3N4P4/c1-13(2)9-14(3,4)11-16(7,8)...\n3021362\tInChI=1S/Cl6F3NSi3/c1-11(2,3)10(12(4,5)7)13(6,8)9\n3342899\tInChI=1S/BrClF4N3P3/c1-10(3)7-11(2,4)9-12(5,6)...\n3654353\tInChI=1S/Cl4NO5PS2/c1-11(2,5-12(3,6)7)10-13(4,8)9\n4466221\tInChI=1S/Cl7N2P3S/c1-10(2,3)8-11(4,5)9-12(6,7)13\n4612505\tInChI=1S/Cl10N4O2P4S/c1-17(2,3)11-19(7,8)13-21...\n5471325\tInChI=1S/Cl2O4Si4/c1-10(2)5-8-3-7-4-9-6-10\n6141923\tInChI=1S/Cl3IO12/c5-1(6,7)14-4(15-2(8,9)10)16-...\n6443167\tInChI=1S/Cl7NSi2/c1-8(9(2,3)4)10(5,6)7\n6731251\tInChI=1S/Cl11N4OP5/c1-17(2,3)12-18(4,5)13-19(6...\n7242878\tInChI=1S/ClFHNO4S2/c1-8(4,5)3-9(2,6)7/h3H\n7310492\tInChI=1S/ClFO5S2/c1-8(3,4)7-9(2,5)6\n8129600\tInChI=1S/Cl12O8Si7/c1-21(2)13-22(3,4)16-26(11)...\n8149394\tInChI=1S/Cl4NO2PS/c1-8(2,3)5-9(4,6)7\n9060183\tInChI=1S/Cl6NSi2/c1-7(8(2)3)9(4,5)6\n9274099\tInChI=1S/Cl3H2N4O2PS2/c1-10(4)5-11(2,8)7-12(3,...\n9377683\tInChI=1S/ClF3NO2PS/c1-9(6,7)5-8(2,3)4\n9447617\tInChI=1S/Cl6N3P3/c1-8-10(3)7-12(5,6)9(2)11(8)4\n9570835\tInChI=1S/Cl3N3O3S3/c1-10(7)4-11(2,8)6-12(3,9)5-10\n9582304\tInChI=1S/Br4Cl2N3P3/c1-10(2)7-11(3,5)9-12(4,6)...\n9852973\tInChI=1S/ClHO4/c2-1(3,4)5/h(H,2,3,4,5)/i2+2,3+...\n```",
    "1267723": "You are absolutely free to ignore as much as the training data as you see fit!",
    "1267726": "Thank you!",
    "1267735": "My concern was that a model that is only able to predict organic molecules will be discarded regardless of the LB score because it must be able to work on those too in the real world where it will be used.\n\nBut **as I understand** now, this is not required and **only LB matters** (as long as all rules are met).",
    "1267797": "Till now, I wrongly called \"carbon-free\" molecules as equal to \"non-organic\" ones."
  },
  "source": "meta"
}