![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F9654628%2F60f271518f38de0c0d00e0607383f5a0%2F3LC-Dashboard-Rendering-Performance.png?generation=1763846925056210&alt=media\")
Hi, I wanted to confirm whether Knowledge Distillation (KD) is allowed in this competition.\nSpecifically, if we train a larger YOLO model (e.g., YOLOv8x) separately and then use it only to distill knowledge into the allowed YOLOv8n model — without submitting or using the larger model directly — would that still comply with the competition rules?
", "messages": [ { "id": 3309798, "postDate": "2025-11-01T12:34:22.117Z", "content": "As long as you end up with YOLOv8n weights that produce your submitted CSV file on inference, it's ok.
\nThe scenario is that you need to deploy YOLOv8n on an edge device (hardware restrictions), which will produce the results in the field. Labels (into the larger YOLOv8x model in your case, presumably) need to be corrected with 3LC. It's a 3LC competition after all 🙂.
\nWhatever happens in between is fair game.
", "rawMarkdown": "As long as you end up with YOLOv8n weights that produce your submitted CSV file on inference, it's ok.\n\nThe scenario is that you need to deploy **YOLOv8n** on an edge device *(hardware restrictions)*, which will **produce the results in the field**. Labels *(into the larger YOLOv8x model in your case, presumably)* need to be corrected with 3LC. *It's a 3LC competition after all 🙂*.\n\nWhatever happens in between is fair game. ", "votes": 1, "replies": [ { "id": 3309816, "postDate": "2025-11-01T13:55:57.723Z", "content": "Side comment:
\nIf you decide to go this route, know that 3LC can handle multiple sets of BBs and Embeddings per sample, and plot/chart them together. So, you can easily capture metrics from both the large and small model inferences, and then use advanced plots to see where the two models differ the most, which samples are easy for the large model but really hard for the small model, and so on. This can give you clues as to which samples should or should not be included, how they should be weighted, and so on.
\nIt may provide you with significant insights into the teacher-student distillation process.
", "rawMarkdown": "Side comment:\n\nIf you decide to go this route, know that 3LC can handle multiple sets of BBs and Embeddings per sample, and plot/chart them together. So, you can easily capture metrics from both the large and small model inferences, and then use advanced plots to see where the two models differ the most, which samples are easy for the large model but really hard for the small model, and so on. This can give you clues as to which samples should or should not be included, how they should be weighted, and so on. \n\nIt may provide you with significant insights into the teacher-student distillation process.", "votes": 1, "replies": [ { "id": 3312195, "postDate": "2025-11-06T17:13:51.383Z", "content": "Hey, can you just clarify whether the training data contains any incorrect bounding boxes or class labels that we’re supposed to detect and correct using 3LC? Also, I’ve tried multiple times to use 3LC by adding the API key and following the notebook tutorial, but I keep getting a “Could not connect to 3LC server” message on the dashboard. I saw on YouTube that we need to run the 3LC service, but in Kaggle notebooks it keeps running indefinitely, so I can’t execute other cells. Could you please confirm the correct way to use 3LC in Kaggle?
", "rawMarkdown": "Hey, can you just clarify whether the training data contains any incorrect bounding boxes or class labels that we’re supposed to detect and correct using 3LC? Also, I’ve tried multiple times to use 3LC by adding the API key and following the notebook tutorial, but I keep getting a “Could not connect to 3LC server” message on the dashboard. I saw on YouTube that we need to run the 3LC service, but in Kaggle notebooks it keeps running indefinitely, so I can’t execute other cells. Could you please confirm the correct way to use 3LC in Kaggle?" }, { "id": 3312248, "postDate": "2025-11-06T19:00:51.703Z", "content": "Yes, data editing is the main part of this challenge. That's also part of the reasons that we have quite a few constraints on the model side, so that we have a fair competition. Please also refer to the \"Data reality\" under the challenge Overview.
", "rawMarkdown": "Yes, data editing is the main part of this challenge. That's also part of the reasons that we have quite a few constraints on the model side, so that we have a fair competition. Please also refer to the \"Data reality\" under the challenge Overview. " }, { "id": 3312268, "postDate": "2025-11-06T19:56:08.167Z", "content": "hey can u give me the clarification on how to use 3cl in kaggle notebook. when i open https://dashboard.3lc.ai/ its showing me could not connect to 3lc server. click here to help running the server.
", "rawMarkdown": "hey can u give me the clarification on how to use 3cl in kaggle notebook. when i open https://dashboard.3lc.ai/ its showing me could not connect to 3lc server. click here to help running the server. " }, { "id": 3312276, "postDate": "2025-11-06T21:11:18.017Z", "content": "The setup provided in the accompanying Notebook details a local setup, where you will start the 3lc service on your local machine to inspect and correct the dataset, and retrain your model.
It is possible to run the training and inference on a Kaggle node (both CPU, GPU, and TPU), but it would require some adaptations, and you would need to download/upload the dataset to Kaggle between each run. If you have the ability to do local training, I would advise to do everything locally. It would be the quickest way to iterate on the dataset.
\n3lc service is the 3LC Object Service which provides all data to the Dashboard.\nThe dashboard you log into at https://dashboard.3lc.ai tries to access your local service on http://localhost:5015
Hey, I am running the 3LC locally and editing and retraining, but I am facing a small issue. I can see the image and the plots by clicking both of them in the 2D plot view. The thing is, it is giving me two 2D plots — one to display the image, and another to display the bounding box with a black background. Is there a way to combine both of them into a single 2D plot (image + bounding box)?
\nBy the way, can you tell me if it's allowed to use data from outside, like using additional weed plant images from external sources on top of what the competition provided?
\nCan you please clarify these two things?
", "rawMarkdown": "Hey, I am running the 3LC locally and editing and retraining, but I am facing a small issue. I can see the image and the plots by clicking both of them in the 2D plot view. The thing is, it is giving me two 2D plots — one to display the image, and another to display the bounding box with a black background. Is there a way to combine both of them into a single 2D plot (image + bounding box)?\n\nBy the way, can you tell me if it's allowed to use data from outside, like using additional weed plant images from external sources on top of what the competition provided?\n\nCan you please clarify these two things?" }, { "id": 3332472, "postDate": "2025-11-16T17:57:45.597Z", "content": "Simply select the two columns you want to plot together and press <2>, and you will get a single view with both image and BBs combined. If you want to plot multiple metrics together, you can even create 3D plots (3 axis), 4D plots (+dot color), and 5D plots (+dot size). So, there's lots of options and tools to figure out which sample is correct/important/redundant, when you plot GT and Predictions together or next to each other.
\nNo, you cannot add outside images in this competition. As this is a simulated business setting, imagine that this is all the data that is available to you, and your budget for further data gathering in the field is zero.
\nDo your best with the data available to you in this challenge.
", "rawMarkdown": "Simply select the two columns you want to plot together and press <2>, and you will get a single view with both image and BBs combined. If you want to plot multiple metrics together, you can even create 3D plots (3 axis), 4D plots (+dot color), and 5D plots (+dot size). So, there's lots of options and tools to figure out which sample is correct/important/redundant, when you plot GT and Predictions together or next to each other.\n\nNo, you cannot add outside images in this competition. As this is a simulated business setting, imagine that this is all the data that is available to you, and your budget for further data gathering in the field is zero.\n\nDo your best with the data available to you in this challenge.", "votes": 1 }, { "id": 3336678, "postDate": "2025-11-18T15:28:17.790Z", "content": "hey i am facing a small problem in dashboard the images are too blur. i can't see the image properly why is that? not network issue for sure its been like that from biggening
", "rawMarkdown": "hey i am facing a small problem in dashboard the images are too blur. i can't see the image properly why is that? not network issue for sure its been like that from biggening" }, { "id": 3344744, "postDate": "2025-11-22T21:28:48.550Z", "content": "This is likely due to rendering settings, and you can change those to fit your system. \n3LC tries to detect the performance of your graphics card, and set settings based on that. But, there are so many systems out there that it can make mistakes, also depending on which browser you are using.
\nIf you press '0' (zero), then you will get the settings panel. Go to the \"Rendering Performance\" and chose settings that work well for you.
\n
Hi, I wanted to confirm whether Knowledge Distillation (KD) is allowed in this competition.\nSpecifically, if we train a larger YOLO model (e.g., YOLOv8x) separately and then use it only to distill knowledge into the allowed YOLOv8n model — without submitting or using the larger model directly — would that still comply with the competition rules?
", "rawMarkdown": "Hi, I wanted to confirm whether Knowledge Distillation (KD) is allowed in this competition.\nSpecifically, if we train a larger YOLO model (e.g., YOLOv8x) separately and then use it only to distill knowledge into the allowed YOLOv8n model — without submitting or using the larger model directly — would that still comply with the competition rules?" }, { "id": 3309796, "postDate": "2025-11-01T12:33:27.920Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 3309798, "author_name": "Marius Storm-Olsen", "author_url": "", "post_date": "2025-11-01T12:34:22.117000", "content": "As long as you end up with YOLOv8n weights that produce your submitted CSV file on inference, it's ok.
\nThe scenario is that you need to deploy YOLOv8n on an edge device (hardware restrictions), which will produce the results in the field. Labels (into the larger YOLOv8x model in your case, presumably) need to be corrected with 3LC. It's a 3LC competition after all 🙂.
\nWhatever happens in between is fair game.
", "votes": 1, "replies": [ { "id": 3309816, "author_name": "Marius Storm-Olsen", "author_url": "", "post_date": "2025-11-01T13:55:57.723000", "content": "Side comment:
\nIf you decide to go this route, know that 3LC can handle multiple sets of BBs and Embeddings per sample, and plot/chart them together. So, you can easily capture metrics from both the large and small model inferences, and then use advanced plots to see where the two models differ the most, which samples are easy for the large model but really hard for the small model, and so on. This can give you clues as to which samples should or should not be included, how they should be weighted, and so on.
\nIt may provide you with significant insights into the teacher-student distillation process.
", "votes": 1, "replies": [ { "id": 3312195, "author_name": "mohanapavan bezawada", "author_url": "", "post_date": "2025-11-06T17:13:51.383000", "content": "Hey, can you just clarify whether the training data contains any incorrect bounding boxes or class labels that we’re supposed to detect and correct using 3LC? Also, I’ve tried multiple times to use 3LC by adding the API key and following the notebook tutorial, but I keep getting a “Could not connect to 3LC server” message on the dashboard. I saw on YouTube that we need to run the 3LC service, but in Kaggle notebooks it keeps running indefinitely, so I can’t execute other cells. Could you please confirm the correct way to use 3LC in Kaggle?
", "votes": 0, "replies": [] }, { "id": 3312248, "author_name": "Dian He", "author_url": "", "post_date": "2025-11-06T19:00:51.703000", "content": "Yes, data editing is the main part of this challenge. That's also part of the reasons that we have quite a few constraints on the model side, so that we have a fair competition. Please also refer to the \"Data reality\" under the challenge Overview.
", "votes": 0, "replies": [] }, { "id": 3312268, "author_name": "mohanapavan bezawada", "author_url": "", "post_date": "2025-11-06T19:56:08.167000", "content": "hey can u give me the clarification on how to use 3cl in kaggle notebook. when i open https://dashboard.3lc.ai/ its showing me could not connect to 3lc server. click here to help running the server.
", "votes": 0, "replies": [] }, { "id": 3312276, "author_name": "Marius Storm-Olsen", "author_url": "", "post_date": "2025-11-06T21:11:18.017000", "content": "The setup provided in the accompanying Notebook details a local setup, where you will start the 3lc service on your local machine to inspect and correct the dataset, and retrain your model.
It is possible to run the training and inference on a Kaggle node (both CPU, GPU, and TPU), but it would require some adaptations, and you would need to download/upload the dataset to Kaggle between each run. If you have the ability to do local training, I would advise to do everything locally. It would be the quickest way to iterate on the dataset.
\n3lc service is the 3LC Object Service which provides all data to the Dashboard.\nThe dashboard you log into at https://dashboard.3lc.ai tries to access your local service on http://localhost:5015
Hey, I am running the 3LC locally and editing and retraining, but I am facing a small issue. I can see the image and the plots by clicking both of them in the 2D plot view. The thing is, it is giving me two 2D plots — one to display the image, and another to display the bounding box with a black background. Is there a way to combine both of them into a single 2D plot (image + bounding box)?
\nBy the way, can you tell me if it's allowed to use data from outside, like using additional weed plant images from external sources on top of what the competition provided?
\nCan you please clarify these two things?
", "votes": 0, "replies": [] }, { "id": 3332472, "author_name": "Marius Storm-Olsen", "author_url": "", "post_date": "2025-11-16T17:57:45.597000", "content": "Simply select the two columns you want to plot together and press <2>, and you will get a single view with both image and BBs combined. If you want to plot multiple metrics together, you can even create 3D plots (3 axis), 4D plots (+dot color), and 5D plots (+dot size). So, there's lots of options and tools to figure out which sample is correct/important/redundant, when you plot GT and Predictions together or next to each other.
\nNo, you cannot add outside images in this competition. As this is a simulated business setting, imagine that this is all the data that is available to you, and your budget for further data gathering in the field is zero.
\nDo your best with the data available to you in this challenge.
", "votes": 1, "replies": [] }, { "id": 3336678, "author_name": "mohanapavan bezawada", "author_url": "", "post_date": "2025-11-18T15:28:17.790000", "content": "hey i am facing a small problem in dashboard the images are too blur. i can't see the image properly why is that? not network issue for sure its been like that from biggening
", "votes": 0, "replies": [] }, { "id": 3344744, "author_name": "Marius Storm-Olsen", "author_url": "", "post_date": "2025-11-22T21:28:48.550000", "content": "This is likely due to rendering settings, and you can change those to fit your system. \n3LC tries to detect the performance of your graphics card, and set settings based on that. But, there are so many systems out there that it can make mistakes, also depending on which browser you are using.
\nIf you press '0' (zero), then you will get the settings panel. Go to the \"Rendering Performance\" and chose settings that work well for you.
\n