{
  "id": 614042,
  "title": "CV and LB and folds that work [Single Model]",
  "url": "/competitions/csiro-biomass/discussion/614042",
  "author_name": "Gaurav Rawat",
  "post_date": "2025-10-31T16:31:46.741000",
  "votes": 17,
  "comment_count": 32,
  "views": 0,
  "content": "<p>Just starting a post to see what correlation people are getting based on the fold strategy whether disclosed or not .</p>\n<p>Basic log for now testing correlation</p>\n<pre><code>- CV 0.30 LB 0.37 with tta  | NN\n- CV 0.54 LB 0.57 with tta | NN\n- CV 0.57 LB 0.51 :( | NN\n- cv 0.598 LB 0.6  | NN\n- cv 0.6 LB 0.61 | NN\n- CV 0.6132 LB 0.60 | Tabular\n- cv 0.6198 LB 0.64 | nn\n- cv 0.6269 LB 0.61 | Tabular\n- cv 0.65 LB 0.62 | Tabular\n- cv 0.7 LB 0.63 | Tabular\n- cv 0.6264 LB 0.65 | NN\n- CV 0.60  0.67 | NN :|\n- CV 0.63 0.7 - NN \n- cv 0.71 0.74 nn\n- cv - 71 76 NN \n</code></pre>",
  "messages": [
    {
      "id": 3309405,
      "postDate": "2025-10-31T16:31:46.743Z",
      "content": "<p>Just starting a post to see what correlation people are getting based on the fold strategy whether disclosed or not .</p>\n<p>Basic log for now testing correlation</p>\n<pre><code>- CV 0.30 LB 0.37 with tta  | NN\n- CV 0.54 LB 0.57 with tta | NN\n- CV 0.57 LB 0.51 :( | NN\n- cv 0.598 LB 0.6  | NN\n- cv 0.6 LB 0.61 | NN\n- CV 0.6132 LB 0.60 | Tabular\n- cv 0.6198 LB 0.64 | nn\n- cv 0.6269 LB 0.61 | Tabular\n- cv 0.65 LB 0.62 | Tabular\n- cv 0.7 LB 0.63 | Tabular\n- cv 0.6264 LB 0.65 | NN\n- CV 0.60  0.67 | NN :|\n- CV 0.63 0.7 - NN \n- cv 0.71 0.74 nn\n- cv - 71 76 NN \n</code></pre>",
      "rawMarkdown": "Just starting a post to see what correlation people are getting based on the fold strategy whether disclosed or not .\n\nBasic log for now testing correlation\n\n```python\n- CV 0.30 LB 0.37 with tta  | NN\n- CV 0.54 LB 0.57 with tta | NN\n- CV 0.57 LB 0.51 :( | NN\n- cv 0.598 LB 0.6  | NN\n- cv 0.6 LB 0.61 | NN\n- CV 0.6132 LB 0.60 | Tabular\n- cv 0.6198 LB 0.64 | nn\n- cv 0.6269 LB 0.61 | Tabular\n- cv 0.65 LB 0.62 | Tabular\n- cv 0.7 LB 0.63 | Tabular\n- cv 0.6264 LB 0.65 | NN\n- CV 0.60  0.67 | NN :|\n- CV 0.63 0.7 - NN \n- cv 0.71 0.74 nn\n- cv - 71 76 NN \n```",
      "votes": 17
    },
    {
      "id": 3309605,
      "postDate": "2025-11-01T03:16:16.093Z",
      "content": "<p>CV: 0.66, LB: 0.69</p>",
      "rawMarkdown": "CV: 0.66, LB: 0.69",
      "votes": 7,
      "replies": [
        {
          "id": 3309809,
          "postDate": "2025-11-01T13:23:59.770Z",
          "content": "<p>Great what fold strategy you used if can disclose .</p>",
          "rawMarkdown": "Great what fold strategy you used if can disclose .",
          "votes": 2,
          "replies": [
            {
              "id": 3310538,
              "postDate": "2025-11-03T06:20:01.037Z",
              "content": "<p>Same as others. I changed evaluation metric according to the recent post, but the gap got worse. I need to check out the alignment.</p>\n<p>CV: 0.79, LB: 0.71</p>",
              "rawMarkdown": "Same as others. I changed evaluation metric according to the recent post, but the gap got worse. I need to check out the alignment.\n\nCV: 0.79, LB: 0.71",
              "votes": 3
            },
            {
              "id": 3313229,
              "postDate": "2025-11-09T00:46:04.197Z",
              "content": "<p>great are u freezing 2 stage or is it just single run </p>",
              "rawMarkdown": "great are u freezing 2 stage or is it just single run "
            }
          ]
        }
      ]
    },
    {
      "id": 3315619,
      "postDate": "2025-11-10T17:31:41.960Z",
      "content": "<p>CV: 0.77, LB: 0.71</p>",
      "rawMarkdown": "CV: 0.77, LB: 0.71",
      "votes": 5,
      "replies": [
        {
          "id": 3318806,
          "postDate": "2025-11-11T12:09:13.730Z",
          "content": "<p>Nice single model assuming</p>",
          "rawMarkdown": "Nice single model assuming",
          "replies": [
            {
              "id": 3350949,
              "postDate": "2025-11-28T02:37:40.710Z",
              "content": "<p>Dunno got now 0.63 - 0.7 the correlation seeing across folks seems confusing 🫤 </p>",
              "rawMarkdown": "Dunno got now 0.63 - 0.7 the correlation seeing across folks seems confusing 🫤 "
            },
            {
              "id": 3351369,
              "postDate": "2025-11-28T11:10:43.970Z",
              "content": "<p>For me, i'm always getting better score in cv than lb using stratifiedgroupkfold. Can you disclose the validation method you are using?</p>",
              "rawMarkdown": "For me, i'm always getting better score in cv than lb using stratifiedgroupkfold. Can you disclose the validation method you are using?"
            }
          ]
        }
      ]
    },
    {
      "id": 3353287,
      "postDate": "2025-11-29T16:49:25.207Z",
      "content": "<p>CV: 0.68, LB: 0.68</p>\n<p>CV: 0.69, LB: 0.69</p>\n<p>CV: 0.72, LB: 0.70 (Some architectural changes)</p>",
      "rawMarkdown": "CV: 0.68, LB: 0.68\n\nCV: 0.69, LB: 0.69\n\nCV: 0.72, LB: 0.70 (Some architectural changes)",
      "votes": 2
    },
    {
      "id": 3311226,
      "postDate": "2025-11-04T14:27:07.413Z",
      "content": "<p><a href=\"https://github.com/bogoconic1/mle-bench/blob/main/mlebench/competitions/csiro-biomass/prepare.py\" target=\"_blank\">Holdout 10% (train-test-split by image_id)</a></p>\n<p>to be updated</p>\n<table>\n<thead>\n<tr>\n<th>CV</th>\n<th>LB</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>0.33065</td>\n<td>0.29</td>\n</tr>\n<tr>\n<td>0.46218</td>\n<td>0.48</td>\n</tr>\n<tr>\n<td>0.54143</td>\n<td>0.56</td>\n</tr>\n</tbody>\n</table>",
      "rawMarkdown": "[Holdout 10% (train-test-split by image_id)](https://github.com/bogoconic1/mle-bench/blob/main/mlebench/competitions/csiro-biomass/prepare.py)\n\nto be updated\n\n| CV | LB |\n| --- | --- |\n| 0.33065  | 0.29 |\n| 0.46218 | 0.48 | \n| 0.54143 | 0.56 |\n\n",
      "votes": 1,
      "replies": [
        {
          "id": 3313228,
          "postDate": "2025-11-09T00:45:34.597Z",
          "content": "<p>dual stage or single stage wihtout freezing ?</p>",
          "rawMarkdown": "dual stage or single stage wihtout freezing ?"
        }
      ]
    },
    {
      "id": 3309639,
      "postDate": "2025-11-01T05:02:14.287Z",
      "content": "<p>Well, even though the size of dataset is small, my CV and LB are well aligned, I tested another submission with CV 0.54 and LB was 0.54 too!</p>",
      "rawMarkdown": "Well, even though the size of dataset is small, my CV and LB are well aligned, I tested another submission with CV 0.54 and LB was 0.54 too!",
      "votes": 1,
      "replies": [
        {
          "id": 3309647,
          "postDate": "2025-11-01T05:12:34.413Z",
          "content": "<p>How did you split your data? I think using date and state could work, but I haven’t tried it yet.</p>",
          "rawMarkdown": "How did you split your data? I think using date and state could work, but I haven’t tried it yet.",
          "replies": [
            {
              "id": 3309650,
              "postDate": "2025-11-01T05:41:19.143Z",
              "content": "<p>K Fold with shuffle, I am not using any metadata</p>",
              "rawMarkdown": "K Fold with shuffle, I am not using any metadata"
            }
          ]
        }
      ]
    },
    {
      "id": 3309458,
      "postDate": "2025-10-31T18:37:03.360Z",
      "content": "<p>got 0.74 on CV</p>\n<p>but 0.58 on LB 😭</p>",
      "rawMarkdown": "got 0.74 on CV\n\nbut 0.58 on LB 😭",
      "votes": 1,
      "replies": [
        {
          "id": 3309485,
          "postDate": "2025-10-31T20:18:29.163Z",
          "content": "<p>cv seems nice</p>",
          "rawMarkdown": "cv seems nice"
        },
        {
          "id": 3369933,
          "postDate": "2025-12-10T13:10:16.807Z",
          "content": "<p>How you fixed the gap?</p>",
          "rawMarkdown": "How you fixed the gap?",
          "replies": [
            {
              "id": 3370043,
              "postDate": "2025-12-10T14:35:11.920Z",
              "content": "<p>That is of course the $50,000 question :)</p>",
              "rawMarkdown": "That is of course the $50,000 question :)",
              "votes": 2
            }
          ]
        }
      ]
    },
    {
      "id": 3309411,
      "postDate": "2025-10-31T16:42:21.760Z",
      "content": "<p>GroupKFold grouping State + Sampling_Date ?</p>",
      "rawMarkdown": "GroupKFold grouping State + Sampling_Date ?",
      "votes": 1,
      "replies": [
        {
          "id": 3309484,
          "postDate": "2025-10-31T20:18:17.210Z",
          "content": "<p>yes trying that so far </p>",
          "rawMarkdown": "yes trying that so far "
        }
      ]
    },
    {
      "id": 3384419,
      "postDate": "2026-01-01T08:12:41.533Z",
      "content": "<p>I'm rookie in cv competition,it seems really hard to improve when score comes to around 0.73-0.74,i have tried a lot of strategies and small changes of architectures,however,all changes didn't improve my local cv score.But maybe i am too stubborn or wrong?Maybe i should try to submit more even  local cv didn't improve?The correlation when score comes to a bottleneck seems a little bit strange though.I will be really glad if you can give me some advice.</p>",
      "rawMarkdown": "I'm rookie in cv competition,it seems really hard to improve when score comes to around 0.73-0.74,i have tried a lot of strategies and small changes of architectures,however,all changes didn't improve my local cv score.But maybe i am too stubborn or wrong?Maybe i should try to submit more even  local cv didn't improve?The correlation when score comes to a bottleneck seems a little bit strange though.I will be really glad if you can give me some advice.",
      "replies": [
        {
          "id": 3386118,
          "postDate": "2026-01-04T16:36:38.403Z",
          "content": "<p>Your doing well for rookie.  This is a horrible data set with a ton of risks that the final test images can make a mess out of models.  </p>\n<p>You get two submission choices for the final scoring.  I like to pick a solid model supported by my local cv result and a much more risky model that might include a bunch of tricks known to sometimes work.  </p>\n<p>In my first competition I had a ton of knowledge about the topic and had a winning model (top 20) in the first 3 weeks.  But I went full blast and ignored local cv for leader board improvement.  I think I ended up in 3198th place after 3 months.  It was a painful lesson on over fitting.</p>",
          "rawMarkdown": "Your doing well for rookie.  This is a horrible data set with a ton of risks that the final test images can make a mess out of models.  \n\nYou get two submission choices for the final scoring.  I like to pick a solid model supported by my local cv result and a much more risky model that might include a bunch of tricks known to sometimes work.  \n\nIn my first competition I had a ton of knowledge about the topic and had a winning model (top 20) in the first 3 weeks.  But I went full blast and ignored local cv for leader board improvement.  I think I ended up in 3198th place after 3 months.  It was a painful lesson on over fitting."
        }
      ]
    },
    {
      "id": 3383777,
      "postDate": "2025-12-30T20:18:48.090Z",
      "content": "<p>are training features like height or NDVI will be available in testing data ?</p>",
      "rawMarkdown": "are training features like height or NDVI will be available in testing data ?",
      "replies": [
        {
          "id": 3383816,
          "postDate": "2025-12-30T22:11:36.967Z",
          "content": "<p>Not available.   If you want to use them you need to build a model to predict them or use median's (or mean) when predicting.   Just forked a <a href=\"https://www.kaggle.com/code/umerhaddii/cisro-fold\" target=\"_blank\">notebook</a> that did this shared by a grand master.   Give it a look.</p>",
          "rawMarkdown": "Not available.   If you want to use them you need to build a model to predict them or use median's (or mean) when predicting.   Just forked a [notebook](https://www.kaggle.com/code/umerhaddii/cisro-fold) that did this shared by a grand master.   Give it a look.",
          "votes": 1
        }
      ]
    },
    {
      "id": 3383436,
      "postDate": "2025-12-30T06:15:05.393Z",
      "content": "<p>Seems that without tta is better.</p>",
      "rawMarkdown": "Seems that without tta is better."
    },
    {
      "id": 3380994,
      "postDate": "2025-12-23T14:41:14.897Z",
      "content": "<p>CV 0.706 LB 0.72</p>\n<p>Update: CV 0.717 LB 0.73</p>\n<p>Update 2: CV 0.711 LB 0.76</p>\n<p>Update 3: AI agent becomes completely useless once it reaches mid 0.76s. It kept going in circles since then.</p>",
      "rawMarkdown": "CV 0.706 LB 0.72\n\nUpdate: CV 0.717 LB 0.73\n\nUpdate 2: CV 0.711 LB 0.76\n\nUpdate 3: AI agent becomes completely useless once it reaches mid 0.76s. It kept going in circles since then.",
      "replies": [
        {
          "id": 3384454,
          "postDate": "2026-01-01T09:14:41.480Z",
          "content": "<p>with this split, we had multiple runs of &gt; 0.75 CV but none of them had LB &gt; 0.7 🤣</p>",
          "rawMarkdown": "with this split, we had multiple runs of > 0.75 CV but none of them had LB > 0.7 🤣",
          "votes": 1
        }
      ]
    },
    {
      "id": 3380054,
      "postDate": "2025-12-21T14:24:58.077Z",
      "content": "<p>Has anyone got like a 0.76+ single model</p>",
      "rawMarkdown": "Has anyone got like a 0.76+ single model"
    },
    {
      "id": 3309752,
      "postDate": "2025-11-01T10:13:05.017Z",
      "content": "<p>CV: 0.62, LB: 0.56</p>\n<p>CV: 0.74, LB: 0.59</p>\n<p>CV: 0.71, LB: <strong><em>scoring</em></strong></p>",
      "rawMarkdown": "CV: 0.62, LB: 0.56\n\nCV: 0.74, LB: 0.59\n\nCV: 0.71, LB: ***scoring***"
    },
    {
      "id": 3309630,
      "postDate": "2025-11-01T04:38:40.510Z",
      "content": "<p>CV: 0.5736, LB:0.58</p>",
      "rawMarkdown": "CV: 0.5736, LB:0.58"
    },
    {
      "id": 3386012,
      "postDate": "2026-01-04T11:40:34.320Z",
      "rawMarkdown": "",
      "isDeleted": true
    },
    {
      "id": 3321017,
      "postDate": "2025-11-13T01:37:11.107Z",
      "rawMarkdown": "",
      "votes": -1,
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 3309605,
      "author_name": "Jungwoo Park",
      "author_url": "",
      "post_date": "2025-11-01T03:16:16.093000",
      "content": "<p>CV: 0.66, LB: 0.69</p>",
      "votes": 7,
      "replies": [
        {
          "id": 3309809,
          "author_name": "Gaurav Rawat",
          "author_url": "",
          "post_date": "2025-11-01T13:23:59.770000",
          "content": "<p>Great what fold strategy you used if can disclose .</p>",
          "votes": 2,
          "replies": [
            {
              "id": 3310538,
              "author_name": "Jungwoo Park",
              "author_url": "",
              "post_date": "2025-11-03T06:20:01.037000",
              "content": "<p>Same as others. I changed evaluation metric according to the recent post, but the gap got worse. I need to check out the alignment.</p>\n<p>CV: 0.79, LB: 0.71</p>",
              "votes": 3,
              "replies": []
            },
            {
              "id": 3313229,
              "author_name": "Gaurav Rawat",
              "author_url": "",
              "post_date": "2025-11-09T00:46:04.197000",
              "content": "<p>great are u freezing 2 stage or is it just single run </p>",
              "votes": 0,
              "replies": []
            }
          ]
        }
      ]
    },
    {
      "id": 3315619,
      "author_name": "evnsnclr",
      "author_url": "",
      "post_date": "2025-11-10T17:31:41.960000",
      "content": "<p>CV: 0.77, LB: 0.71</p>",
      "votes": 5,
      "replies": [
        {
          "id": 3318806,
          "author_name": "Gaurav Rawat",
          "author_url": "",
          "post_date": "2025-11-11T12:09:13.730000",
          "content": "<p>Nice single model assuming</p>",
          "votes": 0,
          "replies": [
            {
              "id": 3350949,
              "author_name": "Gaurav Rawat",
              "author_url": "",
              "post_date": "2025-11-28T02:37:40.710000",
              "content": "<p>Dunno got now 0.63 - 0.7 the correlation seeing across folks seems confusing 🫤 </p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3351369,
              "author_name": "Nizar Charrada",
              "author_url": "",
              "post_date": "2025-11-28T11:10:43.970000",
              "content": "<p>For me, i'm always getting better score in cv than lb using stratifiedgroupkfold. Can you disclose the validation method you are using?</p>",
              "votes": 0,
              "replies": []
            }
          ]
        }
      ]
    },
    {
      "id": 3353287,
      "author_name": "Komil Parmar",
      "author_url": "",
      "post_date": "2025-11-29T16:49:25.207000",
      "content": "<p>CV: 0.68, LB: 0.68</p>\n<p>CV: 0.69, LB: 0.69</p>\n<p>CV: 0.72, LB: 0.70 (Some architectural changes)</p>",
      "votes": 2,
      "replies": []
    },
    {
      "id": 3311226,
      "author_name": "Geremie Yeo",
      "author_url": "",
      "post_date": "2025-11-04T14:27:07.413000",
      "content": "<p><a href=\"https://github.com/bogoconic1/mle-bench/blob/main/mlebench/competitions/csiro-biomass/prepare.py\" target=\"_blank\">Holdout 10% (train-test-split by image_id)</a></p>\n<p>to be updated</p>\n<table>\n<thead>\n<tr>\n<th>CV</th>\n<th>LB</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>0.33065</td>\n<td>0.29</td>\n</tr>\n<tr>\n<td>0.46218</td>\n<td>0.48</td>\n</tr>\n<tr>\n<td>0.54143</td>\n<td>0.56</td>\n</tr>\n</tbody>\n</table>",
      "votes": 1,
      "replies": [
        {
          "id": 3313228,
          "author_name": "Gaurav Rawat",
          "author_url": "",
          "post_date": "2025-11-09T00:45:34.597000",
          "content": "<p>dual stage or single stage wihtout freezing ?</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 3309639,
      "author_name": "Paritosh Kumar Tripathi",
      "author_url": "",
      "post_date": "2025-11-01T05:02:14.287000",
      "content": "<p>Well, even though the size of dataset is small, my CV and LB are well aligned, I tested another submission with CV 0.54 and LB was 0.54 too!</p>",
      "votes": 1,
      "replies": [
        {
          "id": 3309647,
          "author_name": "none00000",
          "author_url": "",
          "post_date": "2025-11-01T05:12:34.413000",
          "content": "<p>How did you split your data? I think using date and state could work, but I haven’t tried it yet.</p>",
          "votes": 0,
          "replies": [
            {
              "id": 3309650,
              "author_name": "Paritosh Kumar Tripathi",
              "author_url": "",
              "post_date": "2025-11-01T05:41:19.143000",
              "content": "<p>K Fold with shuffle, I am not using any metadata</p>",
              "votes": 0,
              "replies": []
            }
          ]
        }
      ]
    },
    {
      "id": 3309458,
      "author_name": "MD Mushfirat Mohaimin",
      "author_url": "",
      "post_date": "2025-10-31T18:37:03.360000",
      "content": "<p>got 0.74 on CV</p>\n<p>but 0.58 on LB 😭</p>",
      "votes": 1,
      "replies": [
        {
          "id": 3309485,
          "author_name": "Gaurav Rawat",
          "author_url": "",
          "post_date": "2025-10-31T20:18:29.163000",
          "content": "<p>cv seems nice</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 3369933,
          "author_name": "🍀<->🍀",
          "author_url": "",
          "post_date": "2025-12-10T13:10:16.807000",
          "content": "<p>How you fixed the gap?</p>",
          "votes": 0,
          "replies": [
            {
              "id": 3370043,
              "author_name": "PC Jimmmy",
              "author_url": "",
              "post_date": "2025-12-10T14:35:11.920000",
              "content": "<p>That is of course the $50,000 question :)</p>",
              "votes": 2,
              "replies": []
            }
          ]
        }
      ]
    },
    {
      "id": 3309411,
      "author_name": "Sangram Patil",
      "author_url": "",
      "post_date": "2025-10-31T16:42:21.760000",
      "content": "<p>GroupKFold grouping State + Sampling_Date ?</p>",
      "votes": 1,
      "replies": [
        {
          "id": 3309484,
          "author_name": "Gaurav Rawat",
          "author_url": "",
          "post_date": "2025-10-31T20:18:17.210000",
          "content": "<p>yes trying that so far </p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 3384419,
      "author_name": "Bruce Wu",
      "author_url": "",
      "post_date": "2026-01-01T08:12:41.533000",
      "content": "<p>I'm rookie in cv competition,it seems really hard to improve when score comes to around 0.73-0.74,i have tried a lot of strategies and small changes of architectures,however,all changes didn't improve my local cv score.But maybe i am too stubborn or wrong?Maybe i should try to submit more even  local cv didn't improve?The correlation when score comes to a bottleneck seems a little bit strange though.I will be really glad if you can give me some advice.</p>",
      "votes": 0,
      "replies": [
        {
          "id": 3386118,
          "author_name": "PC Jimmmy",
          "author_url": "",
          "post_date": "2026-01-04T16:36:38.403000",
          "content": "<p>Your doing well for rookie.  This is a horrible data set with a ton of risks that the final test images can make a mess out of models.  </p>\n<p>You get two submission choices for the final scoring.  I like to pick a solid model supported by my local cv result and a much more risky model that might include a bunch of tricks known to sometimes work.  </p>\n<p>In my first competition I had a ton of knowledge about the topic and had a winning model (top 20) in the first 3 weeks.  But I went full blast and ignored local cv for leader board improvement.  I think I ended up in 3198th place after 3 months.  It was a painful lesson on over fitting.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 3383777,
      "author_name": "Hassan Razzaq",
      "author_url": "",
      "post_date": "2025-12-30T20:18:48.090000",
      "content": "<p>are training features like height or NDVI will be available in testing data ?</p>",
      "votes": 0,
      "replies": [
        {
          "id": 3383816,
          "author_name": "PC Jimmmy",
          "author_url": "",
          "post_date": "2025-12-30T22:11:36.967000",
          "content": "<p>Not available.   If you want to use them you need to build a model to predict them or use median's (or mean) when predicting.   Just forked a <a href=\"https://www.kaggle.com/code/umerhaddii/cisro-fold\" target=\"_blank\">notebook</a> that did this shared by a grand master.   Give it a look.</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 3383436,
      "author_name": "Suda",
      "author_url": "",
      "post_date": "2025-12-30T06:15:05.393000",
      "content": "<p>Seems that without tta is better.</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 3380994,
      "author_name": "Geremie Yeo",
      "author_url": "",
      "post_date": "2025-12-23T14:41:14.897000",
      "content": "<p>CV 0.706 LB 0.72</p>\n<p>Update: CV 0.717 LB 0.73</p>\n<p>Update 2: CV 0.711 LB 0.76</p>\n<p>Update 3: AI agent becomes completely useless once it reaches mid 0.76s. It kept going in circles since then.</p>",
      "votes": 0,
      "replies": [
        {
          "id": 3384454,
          "author_name": "Geremie Yeo",
          "author_url": "",
          "post_date": "2026-01-01T09:14:41.480000",
          "content": "<p>with this split, we had multiple runs of &gt; 0.75 CV but none of them had LB &gt; 0.7 🤣</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 3380054,
      "author_name": "Gaurav Rawat",
      "author_url": "",
      "post_date": "2025-12-21T14:24:58.077000",
      "content": "<p>Has anyone got like a 0.76+ single model</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 3309752,
      "author_name": "Yehoshua",
      "author_url": "",
      "post_date": "2025-11-01T10:13:05.017000",
      "content": "<p>CV: 0.62, LB: 0.56</p>\n<p>CV: 0.74, LB: 0.59</p>\n<p>CV: 0.71, LB: <strong><em>scoring</em></strong></p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 3309630,
      "author_name": "Paritosh Kumar Tripathi",
      "author_url": "",
      "post_date": "2025-11-01T04:38:40.510000",
      "content": "<p>CV: 0.5736, LB:0.58</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 3386012,
      "author_name": "",
      "author_url": "",
      "post_date": "2026-01-04T11:40:34.320000",
      "content": "",
      "votes": 0,
      "replies": []
    },
    {
      "id": 3321017,
      "author_name": "",
      "author_url": "",
      "post_date": "2025-11-13T01:37:11.107000",
      "content": "",
      "votes": -1,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3309405": "Just starting a post to see what correlation people are getting based on the fold strategy whether disclosed or not .\n\nBasic log for now testing correlation\n\n```python\n- CV 0.30 LB 0.37 with tta  | NN\n- CV 0.54 LB 0.57 with tta | NN\n- CV 0.57 LB 0.51 :( | NN\n- cv 0.598 LB 0.6  | NN\n- cv 0.6 LB 0.61 | NN\n- CV 0.6132 LB 0.60 | Tabular\n- cv 0.6198 LB 0.64 | nn\n- cv 0.6269 LB 0.61 | Tabular\n- cv 0.65 LB 0.62 | Tabular\n- cv 0.7 LB 0.63 | Tabular\n- cv 0.6264 LB 0.65 | NN\n- CV 0.60  0.67 | NN :|\n- CV 0.63 0.7 - NN \n- cv 0.71 0.74 nn\n- cv - 71 76 NN \n```",
    "3309605": "CV: 0.66, LB: 0.69",
    "3315619": "CV: 0.77, LB: 0.71",
    "3353287": "CV: 0.68, LB: 0.68\n\nCV: 0.69, LB: 0.69\n\nCV: 0.72, LB: 0.70 (Some architectural changes)",
    "3311226": "[Holdout 10% (train-test-split by image_id)](https://github.com/bogoconic1/mle-bench/blob/main/mlebench/competitions/csiro-biomass/prepare.py)\n\nto be updated\n\n| CV | LB |\n| --- | --- |\n| 0.33065  | 0.29 |\n| 0.46218 | 0.48 | \n| 0.54143 | 0.56 |\n\n",
    "3309639": "Well, even though the size of dataset is small, my CV and LB are well aligned, I tested another submission with CV 0.54 and LB was 0.54 too!",
    "3309458": "got 0.74 on CV\n\nbut 0.58 on LB 😭",
    "3309411": "GroupKFold grouping State + Sampling_Date ?",
    "3384419": "I'm rookie in cv competition,it seems really hard to improve when score comes to around 0.73-0.74,i have tried a lot of strategies and small changes of architectures,however,all changes didn't improve my local cv score.But maybe i am too stubborn or wrong?Maybe i should try to submit more even  local cv didn't improve?The correlation when score comes to a bottleneck seems a little bit strange though.I will be really glad if you can give me some advice.",
    "3383777": "are training features like height or NDVI will be available in testing data ?",
    "3383436": "Seems that without tta is better.",
    "3380994": "CV 0.706 LB 0.72\n\nUpdate: CV 0.717 LB 0.73\n\nUpdate 2: CV 0.711 LB 0.76\n\nUpdate 3: AI agent becomes completely useless once it reaches mid 0.76s. It kept going in circles since then.",
    "3380054": "Has anyone got like a 0.76+ single model",
    "3309752": "CV: 0.62, LB: 0.56\n\nCV: 0.74, LB: 0.59\n\nCV: 0.71, LB: ***scoring***",
    "3309630": "CV: 0.5736, LB:0.58",
    "3386012": "",
    "3321017": ""
  }
}