I'm looking for 1-2 two teammates from the Gold/Silver Zone.
\nFollowing is my current setup.
: ai4bharat/indicwav2vec_v1_bengali\n : \n : .\n : % Train + Valid + Fleurs + openslr + small portion of common voice\nI still didn't compute the effects of including other datasets (Fleurs + openslr 37 + tiny amount of common voice). I'll share it in a few days.
\nI hope this is useful.
", "messages": [ { "id": "2453101", "postDate": "09/23/2023 19:57:08", "content": "I'm looking for 1-2 two teammates from the Gold/Silver Zone.
\nFollowing is my current setup.
: ai4bharat/indicwav2vec_v1_bengali\n : \n : .\n : % Train + Valid + Fleurs + openslr + small portion of common voice\nI still didn't compute the effects of including other datasets (Fleurs + openslr 37 + tiny amount of common voice). I'll share it in a few days.
\nI hope this is useful.
", "rawMarkdown": "I'm looking for 1-2 two teammates from the Gold/Silver Zone. \nFollowing is my current setup.\n\n```\nModel : ai4bharat/indicwav2vec_v1_bengali\nEpochs : 5\nLB : 0.420\nData : 20% Train + Valid + Fleurs + openslr 37 + small portion of common voice\n```\n\nI still didn't compute the effects of including other datasets (Fleurs + openslr 37 + tiny amount of common voice). I'll share it in a few days.\n\nI hope this is useful.", "votes": null }, { "id": "2453269", "postDate": "09/23/2023 23:06:50", "content": "Do you include all these data into training step by step or all at once. Personally, I have used 40% of competition data once, and the result is disappointing.
", "rawMarkdown": "Do you include all these data into training step by step or all at once. Personally, I have used 40% of competition data once, and the result is disappointing.", "votes": null }, { "id": "2453522", "postDate": "09/24/2023 05:39:06", "content": "All at once and then make a random split. I filtered the training data and used 20% of the filtered data. I will update it
", "rawMarkdown": "All at once and then make a random split. I filtered the training data and used 20% of the filtered data. I will update it", "votes": null }, { "id": "2453625", "postDate": "09/24/2023 07:43:21", "content": "Hi Balaji. I want to try unsupervised training for wav2vec model on unlabeled data. If you are willing, we can cooperate ;)
", "rawMarkdown": "Hi Balaji. I want to try unsupervised training for wav2vec model on unlabeled data. If you are willing, we can cooperate ;)", "votes": null }, { "id": "2453631", "postDate": "09/24/2023 07:58:05", "content": "I have trained for 10 epochs on some of the data you have used, and it gets a similar score as yours. I will probably give the competition dataset another shot.
\nI think perhaps you can train on more common_voice, it gets me from 0.445 to 0.428. Have you think of some other way to better fit with the OOD
", "rawMarkdown": "I have trained for 10 epochs on some of the data you have used, and it gets a similar score as yours. I will probably give the competition dataset another shot. \n\nI think perhaps you can train on more common_voice, it gets me from 0.445 to 0.428. Have you think of some other way to better fit with the OOD", "votes": null }, { "id": "2454136", "postDate": "09/24/2023 15:44:16", "content": "I am using the common voice mentioned in the below link
\nI am planning to add one or two datasets to handle OOD
https://www.kaggle.com/competitions/bengaliai-speech/discussion/435300
", "rawMarkdown": "I am using the common voice mentioned in the below link\nI am planning to add one or two datasets to handle OOD\n\nhttps://www.kaggle.com/competitions/bengaliai-speech/discussion/435300", "votes": null }, { "id": "2454142", "postDate": "09/24/2023 15:46:02", "content": "Hi HuBERT,
\nMy current GPU setup wont be good enough to try unsupervised training. I am planning to focus on adding few more datasets.
\nTo improve the score, I believe we need to tackle
\nI believe Language Modeling may be the key
", "rawMarkdown": "Hi HuBERT, \n\nMy current GPU setup wont be good enough to try unsupervised training. I am planning to focus on adding few more datasets.\n\nTo improve the score, I believe we need to tackle \n1. Punctuations\n2. OOD words\n3. Spelling error\n\nI believe Language Modeling may be the key", "votes": null }, { "id": "2457523", "postDate": "09/27/2023 02:37:40", "content": "Hello Balaji,
\nI am doing adapter traning on ai4bharat/indicwav2vec_v1_bengali with 100 % data.
\nI have many A100s , and my 10 epoch training is able to finish in one day. But I haven't made a submission yet. Are you interested in let me join in your team?
\n", "rawMarkdown": "Hello Balaji,\nI am doing adapter traning on ai4bharat/indicwav2vec_v1_bengali with 100 % data.\nI have many A100s , and my 10 epoch training is able to finish in one day. But I haven't made a submission yet. Are you interested in let me join in your team?\n\n\n\n> Hi HuBERT, \n> \n> My current GPU setup wont be good enough to try unsupervised training. I am planning to focus on adding few more datasets.\n> \n> To improve the score, I believe we need to tackle \n> 1. Punctuations\n> 2. OOD words\n> 3. Spelling error\n> \n> I believe Language Modeling may be the key", "votes": null }, { "id": "2457861", "postDate": "09/27/2023 07:45:33", "content": "Hi HuBERT,
\nMy current GPU setup wont be good enough to try unsupervised training. I am planning to focus on adding few more datasets.
\nTo improve the score, I believe we need to tackle
\n\n
\n- Punctuations
\n- OOD words
\n- Spelling error
\nI believe Language Modeling may be the key
\n
I would recommend you to submit. Not just myself, but many others will be interested in working with you if you can achieve Silver or Gold zone.
\nAll the best
", "rawMarkdown": "I would recommend you to submit. Not just myself, but many others will be interested in working with you if you can achieve Silver or Gold zone.\n\nAll the best", "votes": null } ], "comments": [ { "id": 2453269, "author_name": "renyiwei", "author_url": "", "post_date": "09/23/2023 23:06:50", "content": "Do you include all these data into training step by step or all at once. Personally, I have used 40% of competition data once, and the result is disappointing.
", "votes": null, "replies": [ { "id": 2453522, "author_name": "dhakshiin1601", "author_url": "", "post_date": "09/24/2023 05:39:06", "content": "All at once and then make a random split. I filtered the training data and used 20% of the filtered data. I will update it
", "votes": null, "replies": [ { "id": 2453631, "author_name": "renyiwei", "author_url": "", "post_date": "09/24/2023 07:58:05", "content": "I have trained for 10 epochs on some of the data you have used, and it gets a similar score as yours. I will probably give the competition dataset another shot.
\nI think perhaps you can train on more common_voice, it gets me from 0.445 to 0.428. Have you think of some other way to better fit with the OOD
", "votes": null, "replies": [ { "id": 2454136, "author_name": "dhakshiin1601", "author_url": "", "post_date": "09/24/2023 15:44:16", "content": "I am using the common voice mentioned in the below link
\nI am planning to add one or two datasets to handle OOD
https://www.kaggle.com/competitions/bengaliai-speech/discussion/435300
", "votes": null, "replies": [] } ] } ] } ] }, { "id": 2453625, "author_name": "hubert101", "author_url": "", "post_date": "09/24/2023 07:43:21", "content": "Hi Balaji. I want to try unsupervised training for wav2vec model on unlabeled data. If you are willing, we can cooperate ;)
", "votes": null, "replies": [ { "id": 2454142, "author_name": "dhakshiin1601", "author_url": "", "post_date": "09/24/2023 15:46:02", "content": "Hi HuBERT,
\nMy current GPU setup wont be good enough to try unsupervised training. I am planning to focus on adding few more datasets.
\nTo improve the score, I believe we need to tackle
\nI believe Language Modeling may be the key
", "votes": null, "replies": [ { "id": 2457523, "author_name": "baibizhe1", "author_url": "", "post_date": "09/27/2023 02:37:40", "content": "Hello Balaji,
\nI am doing adapter traning on ai4bharat/indicwav2vec_v1_bengali with 100 % data.
\nI have many A100s , and my 10 epoch training is able to finish in one day. But I haven't made a submission yet. Are you interested in let me join in your team?
\n", "votes": null, "replies": [ { "id": 2457861, "author_name": "dhakshiin1601", "author_url": "", "post_date": "09/27/2023 07:45:33", "content": "Hi HuBERT,
\nMy current GPU setup wont be good enough to try unsupervised training. I am planning to focus on adding few more datasets.
\nTo improve the score, I believe we need to tackle
\n\n
\n- Punctuations
\n- OOD words
\n- Spelling error
\nI believe Language Modeling may be the key
\n
I would recommend you to submit. Not just myself, but many others will be interested in working with you if you can achieve Silver or Gold zone.
\nAll the best
", "votes": null, "replies": [] } ] } ] } ] } ], "raw_markdown_by_id": { "2453101": "I'm looking for 1-2 two teammates from the Gold/Silver Zone. \nFollowing is my current setup.\n\n```\nModel : ai4bharat/indicwav2vec_v1_bengali\nEpochs : 5\nLB : 0.420\nData : 20% Train + Valid + Fleurs + openslr 37 + small portion of common voice\n```\n\nI still didn't compute the effects of including other datasets (Fleurs + openslr 37 + tiny amount of common voice). I'll share it in a few days.\n\nI hope this is useful.", "2453269": "Do you include all these data into training step by step or all at once. Personally, I have used 40% of competition data once, and the result is disappointing.", "2453522": "All at once and then make a random split. I filtered the training data and used 20% of the filtered data. I will update it", "2453625": "Hi Balaji. I want to try unsupervised training for wav2vec model on unlabeled data. If you are willing, we can cooperate ;)", "2453631": "I have trained for 10 epochs on some of the data you have used, and it gets a similar score as yours. I will probably give the competition dataset another shot. \n\nI think perhaps you can train on more common_voice, it gets me from 0.445 to 0.428. Have you think of some other way to better fit with the OOD", "2454136": "I am using the common voice mentioned in the below link\nI am planning to add one or two datasets to handle OOD\n\nhttps://www.kaggle.com/competitions/bengaliai-speech/discussion/435300", "2454142": "Hi HuBERT, \n\nMy current GPU setup wont be good enough to try unsupervised training. I am planning to focus on adding few more datasets.\n\nTo improve the score, I believe we need to tackle \n1. Punctuations\n2. OOD words\n3. Spelling error\n\nI believe Language Modeling may be the key", "2457523": "Hello Balaji,\nI am doing adapter traning on ai4bharat/indicwav2vec_v1_bengali with 100 % data.\nI have many A100s , and my 10 epoch training is able to finish in one day. But I haven't made a submission yet. Are you interested in let me join in your team?\n\n\n\n> Hi HuBERT, \n> \n> My current GPU setup wont be good enough to try unsupervised training. I am planning to focus on adding few more datasets.\n> \n> To improve the score, I believe we need to tackle \n> 1. Punctuations\n> 2. OOD words\n> 3. Spelling error\n> \n> I believe Language Modeling may be the key", "2457861": "I would recommend you to submit. Not just myself, but many others will be interested in working with you if you can achieve Silver or Gold zone.\n\nAll the best" }, "source": "meta" }