We've entered the final phase of the competition - many GMs and Masters wait till the last month of a competition to throw their hats into the ring. There's an advantage in waiting for all the kinks of the competition (in this case, the re-annotation exercise) to shake out. The rankings look volatile given the unusual skew towards junior Kagglers, and largely solo teams.
\nOn the flip side, there's still no good public baseline that's NOT driven by string matching which makes it tougher for late entrants. And it's not clear to me whether there'll be the usual \"free\" ensemble boost from late team mergers - the nature of the task means that you can't just do a naive averaging of individual models.
\nI continue to be pessimistic about a shake-up at the top of the LB.
\nGood luck to everyone!
", "messages": [ { "id": "1321760", "postDate": "05/24/2021 23:29:41", "content": "We've entered the final phase of the competition - many GMs and Masters wait till the last month of a competition to throw their hats into the ring. There's an advantage in waiting for all the kinks of the competition (in this case, the re-annotation exercise) to shake out. The rankings look volatile given the unusual skew towards junior Kagglers, and largely solo teams.
\nOn the flip side, there's still no good public baseline that's NOT driven by string matching which makes it tougher for late entrants. And it's not clear to me whether there'll be the usual \"free\" ensemble boost from late team mergers - the nature of the task means that you can't just do a naive averaging of individual models.
\nI continue to be pessimistic about a shake-up at the top of the LB.
\nGood luck to everyone!
", "rawMarkdown": "We've entered the final phase of the competition - many GMs and Masters wait till the last month of a competition to throw their hats into the ring. There's an advantage in waiting for all the kinks of the competition (in this case, the re-annotation exercise) to shake out. The rankings look volatile given the unusual skew towards junior Kagglers, and largely solo teams. \n\nOn the flip side, there's still no good public baseline that's NOT driven by string matching which makes it tougher for late entrants. And it's not clear to me whether there'll be the usual \"free\" ensemble boost from late team mergers - the nature of the task means that you can't just do a naive averaging of individual models. \n\nI continue to be pessimistic about a shake-up at the top of the LB. \n\nGood luck to everyone!", "votes": null }, { "id": "1322532", "postDate": "05/25/2021 13:48:30", "content": "I think that the more important thing than 1st place is to enjoy life with colleagues.
", "rawMarkdown": "I think that the more important thing than 1st place is to enjoy life with colleagues.", "votes": null }, { "id": "1322769", "postDate": "05/25/2021 16:55:54", "content": "Hi, I joined this week
\nI hope we can enjoy this competition :)
i think there will be LB shake up once BMS competition is over.. this is going to be different kind of experience not having good public NB which beats the baseline. lets enjoy the competition..
", "rawMarkdown": "i think there will be LB shake up once BMS competition is over.. this is going to be different kind of experience not having good public NB which beats the baseline. lets enjoy the competition..", "votes": null }, { "id": "1323126", "postDate": "05/26/2021 02:50:33", "content": "You can already see a number of GMs and Masters starting to make subs over the past day ^^
", "rawMarkdown": "You can already see a number of GMs and Masters starting to make subs over the past day ^^", "votes": null }, { "id": "1323913", "postDate": "05/26/2021 13:57:32", "content": "Here is my baseline model using Spacy classifier NO MATCHING scores .44 on the LB. I have a similar BERT model which scores about the same. I extract potential datasets using capitalized word sequences and looks for patterns like (XXXX) so find acronyms like (BLSA), but will pull anything in a similar format extracting the words before the acronym to avoid having to try and predict on all sentences.
\nI am sharing this in hopes someone can break the log jam I am experiencing with scores and others will at least share their LB scores without any matching.
\nMy problem is, I think, many datasets are not properly capitalized and/or do not have an acronym (BBS), (BLSA) etc.
\nhttps://www.kaggle.com/mlconsult/44-3-spacy-classifier-no-matching
\nAlso, here is a notebook that downloads government dataset titles from GOV APIs. It pulls down about 103,000 line items which may help people train.
\nhttps://www.kaggle.com/mlconsult/govt-dataset-json-to-df
", "rawMarkdown": "Here is my baseline model using Spacy classifier NO MATCHING scores .44 on the LB. I have a similar BERT model which scores about the same. I extract potential datasets using capitalized word sequences and looks for patterns like (XXXX) so find acronyms like (BLSA), but will pull anything in a similar format extracting the words before the acronym to avoid having to try and predict on all sentences.\n\nI am sharing this in hopes someone can break the log jam I am experiencing with scores and others will at least share their LB scores without any matching.\n\nMy problem is, I think, many datasets are not properly capitalized and/or do not have an acronym (BBS), (BLSA) etc.\n\nhttps://www.kaggle.com/mlconsult/44-3-spacy-classifier-no-matching\n\nAlso, here is a notebook that downloads government dataset titles from GOV APIs. It pulls down about 103,000 line items which may help people train.\n\nhttps://www.kaggle.com/mlconsult/govt-dataset-json-to-df", "votes": null } ], "comments": [ { "id": 1322532, "author_name": "woosungyoon", "author_url": "", "post_date": "05/25/2021 13:48:30", "content": "I think that the more important thing than 1st place is to enjoy life with colleagues.
", "votes": null, "replies": [] }, { "id": 1322769, "author_name": "ryunosukeishizaki", "author_url": "", "post_date": "05/25/2021 16:55:54", "content": "Hi, I joined this week
\nI hope we can enjoy this competition :)
i think there will be LB shake up once BMS competition is over.. this is going to be different kind of experience not having good public NB which beats the baseline. lets enjoy the competition..
", "votes": null, "replies": [] }, { "id": 1323126, "author_name": "leecming", "author_url": "", "post_date": "05/26/2021 02:50:33", "content": "You can already see a number of GMs and Masters starting to make subs over the past day ^^
", "votes": null, "replies": [] }, { "id": 1323913, "author_name": "mlconsult", "author_url": "", "post_date": "05/26/2021 13:57:32", "content": "Here is my baseline model using Spacy classifier NO MATCHING scores .44 on the LB. I have a similar BERT model which scores about the same. I extract potential datasets using capitalized word sequences and looks for patterns like (XXXX) so find acronyms like (BLSA), but will pull anything in a similar format extracting the words before the acronym to avoid having to try and predict on all sentences.
\nI am sharing this in hopes someone can break the log jam I am experiencing with scores and others will at least share their LB scores without any matching.
\nMy problem is, I think, many datasets are not properly capitalized and/or do not have an acronym (BBS), (BLSA) etc.
\nhttps://www.kaggle.com/mlconsult/44-3-spacy-classifier-no-matching
\nAlso, here is a notebook that downloads government dataset titles from GOV APIs. It pulls down about 103,000 line items which may help people train.
\n", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1321760": "We've entered the final phase of the competition - many GMs and Masters wait till the last month of a competition to throw their hats into the ring. There's an advantage in waiting for all the kinks of the competition (in this case, the re-annotation exercise) to shake out. The rankings look volatile given the unusual skew towards junior Kagglers, and largely solo teams. \n\nOn the flip side, there's still no good public baseline that's NOT driven by string matching which makes it tougher for late entrants. And it's not clear to me whether there'll be the usual \"free\" ensemble boost from late team mergers - the nature of the task means that you can't just do a naive averaging of individual models. \n\nI continue to be pessimistic about a shake-up at the top of the LB. \n\nGood luck to everyone!", "1322532": "I think that the more important thing than 1st place is to enjoy life with colleagues.", "1322769": "Hi, I joined this week\nI hope we can enjoy this competition :)", "1323121": "i think there will be LB shake up once BMS competition is over.. this is going to be different kind of experience not having good public NB which beats the baseline. lets enjoy the competition..", "1323126": "You can already see a number of GMs and Masters starting to make subs over the past day ^^", "1323913": "Here is my baseline model using Spacy classifier NO MATCHING scores .44 on the LB. I have a similar BERT model which scores about the same. I extract potential datasets using capitalized word sequences and looks for patterns like (XXXX) so find acronyms like (BLSA), but will pull anything in a similar format extracting the words before the acronym to avoid having to try and predict on all sentences.\n\nI am sharing this in hopes someone can break the log jam I am experiencing with scores and others will at least share their LB scores without any matching.\n\nMy problem is, I think, many datasets are not properly capitalized and/or do not have an acronym (BBS), (BLSA) etc.\n\nhttps://www.kaggle.com/mlconsult/44-3-spacy-classifier-no-matching\n\nAlso, here is a notebook that downloads government dataset titles from GOV APIs. It pulls down about 103,000 line items which may help people train.\n\nhttps://www.kaggle.com/mlconsult/govt-dataset-json-to-df" }, "source": "meta" }