I get no ideas what the application it would be applied to if I develop a model for this project.
\nWhen do we have a code and a corresponding document both in demand to be merged ?
\nEither, one write a notebook with document or write code without document.
Give me some scenario so that I can imagine what impact would be like. Knowing that would make me engaged more in competition.
\nRefer to the overview they gave has two possible cases
\n\n\nthe construction of better data filtering and preprocessing pipelines for model training, or automatic assessments of a notebook's readability.
\n
I get no ideas what the application it would be applied to if I develop a model for this project.
\nWhen do we have a code and a corresponding document both in demand to be merged ?
\nEither, one write a notebook with document or write code without document.
Give me some scenario so that I can imagine what impact would be like. Knowing that would make me engaged more in competition.
\nRefer to the overview they gave has two possible cases
\n\n\nthe construction of better data filtering and preprocessing pipelines for model training, or automatic assessments of a notebook's readability.
\n
Good question, here is my thought:
\nThe task here is sentence ordering, you probably already knew its cousin - the word ordering task being utilized as a pretraining objective in BART. This is exactly what I think this competition should be treated as: a pretraining problem.
\nYou are doing self-supervised learning, the end goal here is not about how good a model can predict which cell of code would follow a cell of markdown, its about how good that model can be used in related downstream tasks. Think about this as a long shot to a new version of CoPilot that won't require the whole internet as training data.
Often data scientists (beginners) write code in a very unorganized way. For example, an idea came up - they check it. Then they change something, and so on. Sometimes even notebook might not run from start to finish after that! And just like that, a button appears that will solve all problems
", "rawMarkdown": "Often data scientists (beginners) write code in a very unorganized way. For example, an idea came up - they check it. Then they change something, and so on. Sometimes even notebook might not run from start to finish after that! And just like that, a button appears that will solve all problems", "votes": null }, { "id": "1791561", "postDate": "05/16/2022 06:10:26", "content": "Great question and already some very nice points mentioned by Khoi.
\nMy two cents:
\ncode cell and markdown cell based on the similarity or entailment model score.Looking forward to others thoughts and comments. Thank you.
", "rawMarkdown": "Great question and already some very nice points mentioned by Khoi.\n\nMy two cents:\n* Currently most of the public solutions uses rank ordering to solve the problem. But we could also use methods like textual similarity or textual entailment to solve this one as well.\n* If we have that capability, then we could create a readability score comparing the `code cell` and `markdown cell` based on the similarity or entailment model score.\n* As a further downstream task, if we use some text generation models to generate documents / comments automatically with code as input, then we can use this model to filter out the best ones from the generated ones.\n\nLooking forward to others thoughts and comments. Thank you.", "votes": null }, { "id": "1798239", "postDate": "05/22/2022 19:48:00", "content": "When we want to export the Jupyter notebook to a python code file and deliver it to other people there are comments left by the notebook to the python code converter. It's tedious to manually remove the comments. Using a model only codes from notebook files can be extracted and delivered.
", "rawMarkdown": "When we want to export the Jupyter notebook to a python code file and deliver it to other people there are comments left by the notebook to the python code converter. It's tedious to manually remove the comments. Using a model only codes from notebook files can be extracted and delivered.", "votes": null }, { "id": "1798819", "postDate": "05/23/2022 11:56:40", "content": "To go one level out, I'd say that this isn't merely to build models, but to build up people who know enough about the SOTA in code-oriented transformer models to create CoPilot-type tools.
", "rawMarkdown": "To go one level out, I'd say that this isn't merely to build models, but to build up *people* who know enough about the SOTA in code-oriented transformer models to create CoPilot-type tools.", "votes": null }, { "id": "1811579", "postDate": "06/04/2022 20:50:42", "content": "That doesn't seem very useful. Besides, looking at the current best submissions, I highly doubt that the program will beat any human that knows some data science.
", "rawMarkdown": "That doesn't seem very useful. Besides, looking at the current best submissions, I highly doubt that the program will beat any human that knows some data science.", "votes": null }, { "id": "1811740", "postDate": "06/05/2022 04:24:18", "content": "It was an attempt to joke) Seriously speaking, it is quite possible today (in my opinion) to create a tool that will do something similar to the title picture of the competition. For example, by voice command: \"Next, apply min/max normalization to dataframe\" - insert the desired code snippet.
", "rawMarkdown": "It was an attempt to joke) Seriously speaking, it is quite possible today (in my opinion) to create a tool that will do something similar to the title picture of the competition. For example, by voice command: \"*Next, apply min/max normalization to dataframe*\" - insert the desired code snippet.", "votes": null }, { "id": "1811763", "postDate": "06/05/2022 05:41:38", "content": "Sorry. Sarcasm over text is hard to tell, and I'm no expert at any of this. I don't find it likely that the models themselves will have much applications, though. Deep models will likely only learn a few (or hundreds+) of tricks that help order the code snippets. It can do all that without even knowing any data science terminology, including min/max normalization. For example, chess engines were extremely talented at chess, but not much else.
", "rawMarkdown": "Sorry. Sarcasm over text is hard to tell, and I'm no expert at any of this. I don't find it likely that the models themselves will have much applications, though. Deep models will likely only learn a few (or hundreds+) of tricks that help order the code snippets. It can do all that without even knowing any data science terminology, including min/max normalization. For example, chess engines were extremely talented at chess, but not much else.", "votes": null }, { "id": "1822888", "postDate": "06/16/2022 19:23:58", "content": "I can see the commercial value of this model, like, vscode or pycharm provides a link to sort markdown and code automatically based on the text. As we did for indentation of python code automatically.
\nAnother option may be, it could provide automatic markdown insertion based on the code chunks.
Good question, here is my thought:
\nThe task here is sentence ordering, you probably already knew its cousin - the word ordering task being utilized as a pretraining objective in BART. This is exactly what I think this competition should be treated as: a pretraining problem.
\nYou are doing self-supervised learning, the end goal here is not about how good a model can predict which cell of code would follow a cell of markdown, its about how good that model can be used in related downstream tasks. Think about this as a long shot to a new version of CoPilot that won't require the whole internet as training data.
Often data scientists (beginners) write code in a very unorganized way. For example, an idea came up - they check it. Then they change something, and so on. Sometimes even notebook might not run from start to finish after that! And just like that, a button appears that will solve all problems
", "votes": null, "replies": [ { "id": 1811579, "author_name": "solasky", "author_url": "", "post_date": "06/04/2022 20:50:42", "content": "That doesn't seem very useful. Besides, looking at the current best submissions, I highly doubt that the program will beat any human that knows some data science.
", "votes": null, "replies": [] }, { "id": 1811740, "author_name": "sapr3s", "author_url": "", "post_date": "06/05/2022 04:24:18", "content": "It was an attempt to joke) Seriously speaking, it is quite possible today (in my opinion) to create a tool that will do something similar to the title picture of the competition. For example, by voice command: \"Next, apply min/max normalization to dataframe\" - insert the desired code snippet.
", "votes": null, "replies": [] }, { "id": 1811763, "author_name": "solasky", "author_url": "", "post_date": "06/05/2022 05:41:38", "content": "Sorry. Sarcasm over text is hard to tell, and I'm no expert at any of this. I don't find it likely that the models themselves will have much applications, though. Deep models will likely only learn a few (or hundreds+) of tricks that help order the code snippets. It can do all that without even knowing any data science terminology, including min/max normalization. For example, chess engines were extremely talented at chess, but not much else.
", "votes": null, "replies": [] } ] }, { "id": 1791561, "author_name": "sudalairajkumar", "author_url": "", "post_date": "05/16/2022 06:10:26", "content": "Great question and already some very nice points mentioned by Khoi.
\nMy two cents:
\ncode cell and markdown cell based on the similarity or entailment model score.Looking forward to others thoughts and comments. Thank you.
", "votes": null, "replies": [] }, { "id": 1798239, "author_name": "ranamahmud", "author_url": "", "post_date": "05/22/2022 19:48:00", "content": "When we want to export the Jupyter notebook to a python code file and deliver it to other people there are comments left by the notebook to the python code converter. It's tedious to manually remove the comments. Using a model only codes from notebook files can be extracted and delivered.
", "votes": null, "replies": [] }, { "id": 1798819, "author_name": "happycube", "author_url": "", "post_date": "05/23/2022 11:56:40", "content": "To go one level out, I'd say that this isn't merely to build models, but to build up people who know enough about the SOTA in code-oriented transformer models to create CoPilot-type tools.
", "votes": null, "replies": [] }, { "id": 1822888, "author_name": "ashrafr", "author_url": "", "post_date": "06/16/2022 19:23:58", "content": "I can see the commercial value of this model, like, vscode or pycharm provides a link to sort markdown and code automatically based on the text. As we did for indentation of python code automatically.
\nAnother option may be, it could provide automatic markdown insertion based on the code chunks.