s you can see, there's a large concentration of people in the 35.5-36.2 score range. Furthermore, the language preference style of Akkadian's translation validation set is highly uncertain and may even deviate somewhat. How much do you think the Leaderboard will shift in this competition? Will it be like that infamous Mercedes competition?
", "messages": [ { "id": 3424254, "postDate": "2026-03-19T13:55:16.880Z", "content": "My guess: it will be kind of akin to Make Data Count competition!
", "rawMarkdown": "My guess: it will be kind of akin to Make Data Count competition!", "votes": 1 }, { "id": 3424235, "postDate": "2026-03-19T13:31:05.083Z", "content": "s you can see, there's a large concentration of people in the 35.5-36.2 score range. Furthermore, the language preference style of Akkadian's translation validation set is highly uncertain and may even deviate somewhat. How much do you think the Leaderboard will shift in this competition? Will it be like that infamous Mercedes competition?
", "rawMarkdown": "s you can see, there's a large concentration of people in the 35.5-36.2 score range. Furthermore, the language preference style of Akkadian's translation validation set is highly uncertain and may even deviate somewhat. How much do you think the Leaderboard will shift in this competition? Will it be like that infamous Mercedes competition?", "votes": 1 }, { "id": 3425907, "postDate": "2026-03-21T14:14:50.937Z", "content": "I don't know anymore, with over 60 experiments, my best result was 34.3, which can't compare to the pb kernel :(
", "rawMarkdown": "I don't know anymore, with over 60 experiments, my best result was 34.3, which can't compare to the pb kernel :(", "votes": 2, "replies": [ { "id": 3426070, "postDate": "2026-03-21T20:23:38.657Z", "content": "don't worried,good luck :)
", "rawMarkdown": "don't worried,good luck :)", "votes": 2 } ] }, { "id": 3424339, "postDate": "2026-03-19T15:40:03.503Z", "content": "I think there will be minor shakeup, probably similar to Jigsaw
", "rawMarkdown": "I think there will be minor shakeup, probably similar to Jigsaw", "votes": 2, "replies": [ { "id": 3424818, "postDate": "2026-03-20T01:22:04.630Z", "content": "hope that
", "rawMarkdown": "hope that " } ] }, { "id": 3424243, "postDate": "2026-03-19T13:44:08.740Z", "content": "ICR competition is a better reference - why dont we spend $160k+ on a tabular competition with ~600 samples with some uncertain data.
\n<gap> distribution on the private LB will matter the most I guess. I know my models are bad at predicting good placement of this token.
Have you tried creating a local validation set? I seem to have a hard time figuring out the language style of LB, and I'm afraid that the language style of the remaining data will change, like a lottery.
", "rawMarkdown": "Have you tried creating a local validation set? I seem to have a hard time figuring out the language style of LB, and I'm afraid that the language style of the remaining data will change, like a lottery.\n\n ", "replies": [ { "id": 3425056, "postDate": "2026-03-20T09:20:52.350Z", "content": "All I know for a fact that 35-36 models hallucinate much more than a 40 model (from my exps and my models). I have a holdout set to verify this. Have no idea about a public/private distribution.
\nThere could be some shake because previous tiny data fixes resulted in a huge score discrepancies. But in general 37.0-38.0 << 39.0-40.0.
\nAnything between mid-silver and bottom-bronze will basically be random.shuffle() because they mostly heavily overfit public LB by playing with decoding params and +-0.001 weights in their ensembles. Personally, I just have a single 40.0 model without any decoding tuning
", "rawMarkdown": "All I know for a fact that 35-36 models hallucinate much more than a 40 model (from my exps and my models). I have a holdout set to verify this. Have no idea about a public/private distribution.\n\nThere could be some shake because previous tiny data fixes resulted in a huge score discrepancies. But in general 37.0-38.0 << 39.0-40.0.\n\nAnything between mid-silver and bottom-bronze will basically be random.shuffle() because they mostly heavily overfit public LB by playing with decoding params and +-0.001 weights in their ensembles. Personally, I just have a single 40.0 model without any decoding tuning", "votes": 7 } ] } ] }, { "id": 3424908, "postDate": "2026-03-20T04:39:42.277Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 3424254, "author_name": "ducnh279", "author_url": "", "post_date": "2026-03-19T13:55:16.880000", "content": "My guess: it will be kind of akin to Make Data Count competition!
", "votes": 1, "replies": [] }, { "id": 3425907, "author_name": "datnt114", "author_url": "", "post_date": "2026-03-21T14:14:50.937000", "content": "I don't know anymore, with over 60 experiments, my best result was 34.3, which can't compare to the pb kernel :(
", "votes": 2, "replies": [ { "id": 3426070, "author_name": "耶✌", "author_url": "", "post_date": "2026-03-21T20:23:38.657000", "content": "don't worried,good luck :)
", "votes": 2, "replies": [] } ] }, { "id": 3424339, "author_name": "Geremie Yeo", "author_url": "", "post_date": "2026-03-19T15:40:03.503000", "content": "I think there will be minor shakeup, probably similar to Jigsaw
", "votes": 2, "replies": [ { "id": 3424818, "author_name": "Chen Leonid Shaopeng_bnbu", "author_url": "", "post_date": "2026-03-20T01:22:04.630000", "content": "hope that
", "votes": 0, "replies": [] } ] }, { "id": 3424243, "author_name": "Yurnero", "author_url": "", "post_date": "2026-03-19T13:44:08.740000", "content": "ICR competition is a better reference - why dont we spend $160k+ on a tabular competition with ~600 samples with some uncertain data.
\n<gap> distribution on the private LB will matter the most I guess. I know my models are bad at predicting good placement of this token.
Have you tried creating a local validation set? I seem to have a hard time figuring out the language style of LB, and I'm afraid that the language style of the remaining data will change, like a lottery.
", "votes": 0, "replies": [ { "id": 3425056, "author_name": "Yurnero", "author_url": "", "post_date": "2026-03-20T09:20:52.350000", "content": "All I know for a fact that 35-36 models hallucinate much more than a 40 model (from my exps and my models). I have a holdout set to verify this. Have no idea about a public/private distribution.
\nThere could be some shake because previous tiny data fixes resulted in a huge score discrepancies. But in general 37.0-38.0 << 39.0-40.0.
\nAnything between mid-silver and bottom-bronze will basically be random.shuffle() because they mostly heavily overfit public LB by playing with decoding params and +-0.001 weights in their ensembles. Personally, I just have a single 40.0 model without any decoding tuning
", "votes": 7, "replies": [] } ] } ] }, { "id": 3424908, "author_name": "", "author_url": "", "post_date": "2026-03-20T04:39:42.277000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "3424254": "My guess: it will be kind of akin to Make Data Count competition!", "3424235": "s you can see, there's a large concentration of people in the 35.5-36.2 score range. Furthermore, the language preference style of Akkadian's translation validation set is highly uncertain and may even deviate somewhat. How much do you think the Leaderboard will shift in this competition? Will it be like that infamous Mercedes competition?", "3425907": "I don't know anymore, with over 60 experiments, my best result was 34.3, which can't compare to the pb kernel :(", "3424339": "I think there will be minor shakeup, probably similar to Jigsaw", "3424243": "[ICR competition](https://www.kaggle.com/c/icr-identify-age-related-conditions) is a better reference - why dont we spend $160k+ on a tabular competition with ~600 samples with some uncertain data.\n\n`