by jagofc
The rate of markdown cells in a notebook/number of cells in general looks like a fractal
Flax Code Query embeddings
\nby jdoesv
Language of markdowns
\nby nofreewill
About loss function and better score
\nby yuanzhezhou
Textual Similarity Approach
\nby sudalairajkumar
Proposed some text simillarity approaches to solve this competition:
\nCleaning and Preprocessing Markdown cells
\nby usaiprashanth
Textual Entailment Approach
\nby sudalairajkumar
My understanding of the domain - Taking a step back
\nby allohvk
real world application for this model
\nby superching
Hint for a simple but stronger pairwise model baseline
\nby yuanzhezhou
Shared some ideas and methods for improving any pairwise model:
\nKendall’s Tau (Kendall Rank Correlation Coefficient)
\nby mpwolke
What if you don't have high-end GPU?
\nby locbaop
Pairwise model baseline
\nby yuanzhezhou
Some papers that might help
\nby yuanzhezhou
The code cells are not shuffled
\nby yuanzhezhou
\n\nYour task is to predict the correct ordering of the cells in a given notebook whose markdown cells have been shuffled.
\n
Ideas for Generating New Data 💡
\nby ruchi798
Pretrained Python Models on HuggingFace 🤗
\nby snnclsr
by jagofc
The rate of markdown cells in a notebook/number of cells in general looks like a fractal
Flax Code Query embeddings
\nby jdoesv
Language of markdowns
\nby nofreewill
About loss function and better score
\nby yuanzhezhou
Textual Similarity Approach
\nby sudalairajkumar
Proposed some text simillarity approaches to solve this competition:
\nCleaning and Preprocessing Markdown cells
\nby usaiprashanth
Textual Entailment Approach
\nby sudalairajkumar
My understanding of the domain - Taking a step back
\nby allohvk
real world application for this model
\nby superching
Hint for a simple but stronger pairwise model baseline
\nby yuanzhezhou
Shared some ideas and methods for improving any pairwise model:
\nKendall’s Tau (Kendall Rank Correlation Coefficient)
\nby mpwolke
What if you don't have high-end GPU?
\nby locbaop
Pairwise model baseline
\nby yuanzhezhou
Some papers that might help
\nby yuanzhezhou
The code cells are not shuffled
\nby yuanzhezhou
\n\nYour task is to predict the correct ordering of the cells in a given notebook whose markdown cells have been shuffled.
\n
Ideas for Generating New Data 💡
\nby ruchi798
Pretrained Python Models on HuggingFace 🤗
\nby snnclsr
Proposed 3 methods to generate more data:
\nCode -> Comments
\nComments -> Code
\nExternal data
@thedevastator Valuable Summary!
", "rawMarkdown": "thedevastator Valuable Summary!", "votes": null } ], "comments": [ { "id": 1829605, "author_name": "kanisfatemashanta", "author_url": "", "post_date": "06/22/2022 18:52:16", "content": "Proposed 3 methods to generate more data:
\nCode -> Comments
\nComments -> Code
\nExternal data
@thedevastator Valuable Summary!
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1822248": "#### What do we know so far? - A summary of the topic of the top discussion topics\n\n[Fractal markdown distribution](https://www.kaggle.com/competitions/AI4Code/discussion/327740)\n- by [jagofc](jagofc)\n\n- The rate of markdown cells in a notebook/number of cells in general looks like a fractal\n\n[Flax Code Query embeddings](https://www.kaggle.com/competitions/AI4Code/discussion/324588)\nby [jdoesv](jdoesv)\n\n- [Finding: **Positive**] There's an SBERT model that generates code and query vectors in a neutral space trained. The model is available [here](https://huggingface.co/flax-sentence-embeddings/st-codesearch-distilroberta-base). Anyone who wants to take the semantic search route in this competition will find it quite useful.\n\n[Language of markdowns](https://www.kaggle.com/competitions/AI4Code/discussion/324608)\nby [nofreewill](nofreewill)\n\n- On the final set there will be will be multiple languages represented. Not just english.\n\n[About loss function and better score](https://www.kaggle.com/competitions/AI4Code/discussion/325272)\nby [yuanzhezhou](yuanzhezhou)\n\n- MSE loss and L1 loss functions were compared and discussed in the notebook.\n- MSELoss seems to be not a good loss function for this competition, for the reason that the distribution of predictions made by the model trained by MSELoss is not normally distributed, and far from the distribution of the ground truth labels.\n- L1Loss seems to be a better loss function for this competition, for the reason that the distribution of predictions made by the model trained by L1Loss is similar to the distribution of the ground truth labels.\n\n[Textual Similarity Approach](https://www.kaggle.com/competitions/AI4Code/discussion/324509)\nby [sudalairajkumar](sudalairajkumar)\n\n**Proposed some text simillarity approaches to solve this competition:**\n\n- One possible approach is to use the textual similarity between the code and the markdown cells to identify whether they are next to each other or not.\n- We can rank-order the pairs based on the similarity score and then select the top markdown cell as the next / previous cell.\n- We also need to way to find out whether the markdown cell is the previous or next cell to the give code cell once we find the similarity\n- We also need to have a cutoff if we use this approach since some of the code cells may not have markdown cells next to it.\n- We may also get consecutive markdown cells and so we need to find a way to deal with this scenario as well.\n\n[Cleaning and Preprocessing Markdown cells](https://www.kaggle.com/competitions/AI4Code/discussion/325779)\nby [usaiprashanth](usaiprashanth)\n\n- Found this amazing [gist](https://gist.github.com/lorey/eb15a7f3338f959a78cc3661fbc255fe) on github which recovers plain text from any markdown (or) markdown+html cells\n\n[Textual Entailment Approach](https://www.kaggle.com/competitions/AI4Code/discussion/325560)\nby [sudalairajkumar](sudalairajkumar)\n\n- Proposed yet another another potential approach: [Textual Entailment based methods](https://en.wikipedia.org/wiki/Textual_entailment).\n\n[My understanding of the domain - Taking a step back](https://www.kaggle.com/competitions/AI4Code/discussion/328905)\nby [allohvk](allohvk)\n\n- Discussed the topic in general\n- Learning to Rank framework\n- Code-Aware Language Models\n\n[real world application for this model](https://www.kaggle.com/competitions/AI4Code/discussion/325205)\nby [superching](superching)\n\n- Discussed if the competition had any real value to the real world since when we have both the code and the documentation for the code, well.. we have both.\n\n[Hint for a simple but stronger pairwise model baseline](https://www.kaggle.com/competitions/AI4Code/discussion/327197)\nby [yuanzhezhou](yuanzhezhou)\n\nShared some ideas and methods for improving any pairwise model:\n- Pretraining\n- Sampling strategy\n- Better embedding for cells\n- Better ranking strategy.\n- Global context information\n- Post-processing\n\n[Kendall’s Tau (Kendall Rank Correlation Coefficient)](https://www.kaggle.com/competitions/AI4Code/discussion/324476)\nby [mpwolke](mpwolke)\n\n- Explained the Kendall tau rank correlation metric\n\n[What if you don't have high-end GPU?](https://www.kaggle.com/competitions/AI4Code/discussion/329848)\nby [locbaop](locbaop)\n\n- Discussed options to consider if one doesn't have a GPU.\n- The thread itself is very interesting the different solutions had been proposed by the members of the community.\n\n[Pairwise model baseline](https://www.kaggle.com/competitions/AI4Code/discussion/326473)\nby [yuanzhezhou](yuanzhezhou)\n\n- Shared both a pairwise and pointwise baseline\n\n[Some papers that might help](https://www.kaggle.com/competitions/AI4Code/discussion/325725)\nby [yuanzhezhou](yuanzhezhou)\n\n- Shared 3 different papers about sentence ordering that can be useful for this competition\n\n[The code cells are not shuffled](https://www.kaggle.com/competitions/AI4Code/discussion/324813)\nby [yuanzhezhou](yuanzhezhou)\n\n- Highlighted an important part of the competition's description:\n> Your task is to predict the correct ordering of the cells in a given notebook whose markdown cells have been shuffled.\n\n[Ideas for Generating New Data 💡 ](https://www.kaggle.com/competitions/AI4Code/discussion/324554) \nby [ruchi798](ruchi798)\n\n- Proposed 3 methods to generate more data:\n1. Code -> Comments\n2. Comments -> Code\n3. External data\n\n[Pretrained Python Models on HuggingFace 🤗](https://www.kaggle.com/competitions/AI4Code/discussion/324470) \nby [snnclsr](snnclsr)\n\n- Shared a huge list of pretrained model from huggingface\n\n---", "1829605": "Proposed 3 methods to generate more data:\nCode -> Comments\nComments -> Code\nExternal data", "1829790": "thedevastator Valuable Summary!" }, "source": "meta" }