{ "id": 634176, "title": "Clarification on external pretrained models, public weights & 9h runtime rule", "url": "/competitions/csiro-biomass/discussion/634176", "author_name": "", "post_date": "2025-11-19T13:09:44.014878300Z", "votes": 2, "comment_count": 4, "views": 0, "content": "

Hi, I hope you are doing well.

\n

I’d like to clarify how the rules about external data and runtime apply in the following situation.\nThe Code Requirements say:

\n\n

My questions:

\n

Pretrained weights / “publicly available”

\n

If I train a model, either in a separate Kaggle notebook or offline, and then upload only the resulting weights as a public Kaggle Dataset, is that considered compliant with the “freely & publicly available external data, including pre-trained models” rule?\nOr do you require anything more specific, e.g. explicitly stating in the dataset description that it was trained for this CSIRO competition and listing all data sources used?

\n

Where training happens vs. 9h limit

\n

Am I correct that the 9-hour runtime limit applies only to the notebook that is actually submitted to the competition (the inference / scoring notebook), and not to any separate training notebooks or offline training runs?\nIn other words, is it acceptable that:

\n

Training may take longer than 9 hours, but the submitted notebook only loads the pretrained weights (from a public Kaggle Dataset) and runs inference within the 9-hour limit?

\n

Transparency expectations

\n

For top solutions, do you have any extra expectations regarding how we document our pretrained models (e.g. external data used, training setup) beyond what’s in the standard Kaggle rules?

\n

I just want to make sure our workflow is fully aligned with the competition rules before we commit to a specific setup.

\n

Thanks in advance for the clarification.

", "messages": [ { "id": "3340199", "postDate": "11/19/2025 13:09:44", "content": "

Hi, I hope you are doing well.

\n

I’d like to clarify how the rules about external data and runtime apply in the following situation.\nThe Code Requirements say:

\n\n

My questions:

\n

Pretrained weights / “publicly available”

\n

If I train a model, either in a separate Kaggle notebook or offline, and then upload only the resulting weights as a public Kaggle Dataset, is that considered compliant with the “freely & publicly available external data, including pre-trained models” rule?\nOr do you require anything more specific, e.g. explicitly stating in the dataset description that it was trained for this CSIRO competition and listing all data sources used?

\n

Where training happens vs. 9h limit

\n

Am I correct that the 9-hour runtime limit applies only to the notebook that is actually submitted to the competition (the inference / scoring notebook), and not to any separate training notebooks or offline training runs?\nIn other words, is it acceptable that:

\n

Training may take longer than 9 hours, but the submitted notebook only loads the pretrained weights (from a public Kaggle Dataset) and runs inference within the 9-hour limit?

\n

Transparency expectations

\n

For top solutions, do you have any extra expectations regarding how we document our pretrained models (e.g. external data used, training setup) beyond what’s in the standard Kaggle rules?

\n

I just want to make sure our workflow is fully aligned with the competition rules before we commit to a specific setup.

\n

Thanks in advance for the clarification.

", "rawMarkdown": "Hi, I hope you are doing well.\n\nI’d like to clarify how the rules about external data and runtime apply in the following situation.\nThe Code Requirements say:\n- “Freely & publicly available external data is allowed, including pre-trained models”\n- “CPU/GPU Notebook ≤ 9 hours run-time”\n- “Submissions must be made through Notebooks”\n\n### My questions:\n#### **Pretrained weights / “publicly available”**\nIf I train a model, either in a separate Kaggle notebook or offline, and then upload only the resulting weights as a public Kaggle Dataset, is that considered compliant with the “freely & publicly available external data, including pre-trained models” rule?\nOr do you require anything more specific, e.g. explicitly stating in the dataset description that it was trained for this CSIRO competition and listing all data sources used?\n\n#### **Where training happens vs. 9h limit**\nAm I correct that the 9-hour runtime limit applies only to the notebook that is actually submitted to the competition (the inference / scoring notebook), and not to any separate training notebooks or offline training runs?\nIn other words, is it acceptable that:\n\nTraining may take longer than 9 hours, but the submitted notebook only loads the pretrained weights (from a public Kaggle Dataset) and runs inference within the 9-hour limit?\n\n#### **Transparency expectations**\nFor top solutions, do you have any extra expectations regarding how we document our pretrained models (e.g. external data used, training setup) beyond what’s in the standard Kaggle rules?\n\nI just want to make sure our workflow is fully aligned with the competition rules before we commit to a specific setup.\n\nThanks in advance for the clarification.", "votes": null }, { "id": "3341604", "postDate": "11/20/2025 10:00:20", "content": "

@qiyuliaocsiro Hi, can you please check this questioning when you have time. thank you 🙏

", "rawMarkdown": "qiyuliaocsiro Hi, can you please check this questioning when you have time. thank you 🙏", "votes": null }, { "id": "3342179", "postDate": "11/20/2025 17:14:54", "content": "

Pretty sure most of my answers correct :)

\n

Pretrained weights / “publicly available”

\n

I train almost all my models on local machines since I have 4 of them and almost no life :)\nI load them into datasets for the submission phase in a notebook.\nYou don't need to state anything right now - but if you are Gold and win some cash you will need to provide kaggle with evidence that you complied. Look in the Rules section - it describes.
\nPretty much to win the cash prize you need to prove that you used only data that met the rules and that any pretrained weight models were available to the public. Outside of the money range there's no further proofs required.

\n

Where training happens vs. 9h limit

\n

No big deal if training exceeds 9 but there is normally a limit on run time for any notebook - think its currently 12 hours.\nThis assumes that you train in one notebook and make the submission from another. If you working only on kaggle with no local training its pretty easy to link the saved models in the training notebook to the submission notebook. The submission notebook must meet the 9 hour rule.

\n

Transparency expectations\nThe Rules tab has a link to more of the requirements for winners with regards to documentation

\n

https://www.kaggle.com/WinningModelDocumentationGuidelines

", "rawMarkdown": "Pretty sure most of my answers correct :)\n\n**Pretrained weights / “publicly available”**\n\nI train almost all my models on local machines since I have 4 of them and almost no life :)\nI load them into datasets for the submission phase in a notebook.\nYou don't need to state anything right now - but if you are Gold and win some cash you will need to provide kaggle with evidence that you complied. Look in the Rules section - it describes. \nPretty much to win the cash prize you need to prove that you used only data that met the rules and that any pretrained weight models were available to the public. Outside of the money range there's no further proofs required.\n\n**Where training happens vs. 9h limit**\n\nNo big deal if training exceeds 9 but there is normally a limit on run time for any notebook - think its currently 12 hours.\nThis assumes that you train in one notebook and make the submission from another. If you working only on kaggle with no local training its pretty easy to link the saved models in the training notebook to the submission notebook. The submission notebook must meet the 9 hour rule. \n\n**Transparency expectations**\nThe Rules tab has a link to more of the requirements for winners with regards to documentation\n\nhttps://www.kaggle.com/WinningModelDocumentationGuidelines", "votes": null }, { "id": "3342443", "postDate": "11/21/2025 01:16:03", "content": "

I'm not a Kaggle official but I guess Jimmmy is correct.

", "rawMarkdown": "I'm not a Kaggle official but I guess Jimmmy is correct.", "votes": null }, { "id": "3343386", "postDate": "11/21/2025 15:55:05", "content": "

HI @pcjimmmy , thanks a lot for your answer. So, as far as I understood , I can :

\n\n

N.B : At least you have 4 local machines to train your models, on my side, I only have access to kaggle GPU, and still no life 🥲

", "rawMarkdown": "HI @pcjimmmy , thanks a lot for your answer. So, as far as I understood , I can : \n\n- Use any public pre-trained model.\n- Fine-tune it however I want (local or Kaggle), as long as any extra data is public & allowed.\n- Store fine-tuned checkpoint in a private Kaggle dataset or notebook output.\n- Submission notebook: Attaches that dataset, Loads weights, Runs inference and writes submission.csv within 9h.\n- And more important, I only need to \"prove\" anything if in the prize range, and then the focus is on data sources + reproducibility, not on making tuned weights publicly visible to all competitors.\n\nN.B : At least you have 4 local machines to train your models, on my side, I only have access to kaggle GPU, and still no life 🥲", "votes": null } ], "comments": [ { "id": 3341604, "author_name": "mjdism", "author_url": "", "post_date": "11/20/2025 10:00:20", "content": "

@qiyuliaocsiro Hi, can you please check this questioning when you have time. thank you 🙏

", "votes": null, "replies": [ { "id": 3342443, "author_name": "qiyuliaocsiro", "author_url": "", "post_date": "11/21/2025 01:16:03", "content": "

I'm not a Kaggle official but I guess Jimmmy is correct.

", "votes": null, "replies": [] } ] }, { "id": 3342179, "author_name": "pcjimmmy", "author_url": "", "post_date": "11/20/2025 17:14:54", "content": "

Pretty sure most of my answers correct :)

\n

Pretrained weights / “publicly available”

\n

I train almost all my models on local machines since I have 4 of them and almost no life :)\nI load them into datasets for the submission phase in a notebook.\nYou don't need to state anything right now - but if you are Gold and win some cash you will need to provide kaggle with evidence that you complied. Look in the Rules section - it describes.
\nPretty much to win the cash prize you need to prove that you used only data that met the rules and that any pretrained weight models were available to the public. Outside of the money range there's no further proofs required.

\n

Where training happens vs. 9h limit

\n

No big deal if training exceeds 9 but there is normally a limit on run time for any notebook - think its currently 12 hours.\nThis assumes that you train in one notebook and make the submission from another. If you working only on kaggle with no local training its pretty easy to link the saved models in the training notebook to the submission notebook. The submission notebook must meet the 9 hour rule.

\n

Transparency expectations\nThe Rules tab has a link to more of the requirements for winners with regards to documentation

\n

https://www.kaggle.com/WinningModelDocumentationGuidelines

", "votes": null, "replies": [ { "id": 3343386, "author_name": "mjdism", "author_url": "", "post_date": "11/21/2025 15:55:05", "content": "

HI @pcjimmmy , thanks a lot for your answer. So, as far as I understood , I can :

\n\n

N.B : At least you have 4 local machines to train your models, on my side, I only have access to kaggle GPU, and still no life 🥲

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "3340199": "Hi, I hope you are doing well.\n\nI’d like to clarify how the rules about external data and runtime apply in the following situation.\nThe Code Requirements say:\n- “Freely & publicly available external data is allowed, including pre-trained models”\n- “CPU/GPU Notebook ≤ 9 hours run-time”\n- “Submissions must be made through Notebooks”\n\n### My questions:\n#### **Pretrained weights / “publicly available”**\nIf I train a model, either in a separate Kaggle notebook or offline, and then upload only the resulting weights as a public Kaggle Dataset, is that considered compliant with the “freely & publicly available external data, including pre-trained models” rule?\nOr do you require anything more specific, e.g. explicitly stating in the dataset description that it was trained for this CSIRO competition and listing all data sources used?\n\n#### **Where training happens vs. 9h limit**\nAm I correct that the 9-hour runtime limit applies only to the notebook that is actually submitted to the competition (the inference / scoring notebook), and not to any separate training notebooks or offline training runs?\nIn other words, is it acceptable that:\n\nTraining may take longer than 9 hours, but the submitted notebook only loads the pretrained weights (from a public Kaggle Dataset) and runs inference within the 9-hour limit?\n\n#### **Transparency expectations**\nFor top solutions, do you have any extra expectations regarding how we document our pretrained models (e.g. external data used, training setup) beyond what’s in the standard Kaggle rules?\n\nI just want to make sure our workflow is fully aligned with the competition rules before we commit to a specific setup.\n\nThanks in advance for the clarification.", "3341604": "qiyuliaocsiro Hi, can you please check this questioning when you have time. thank you 🙏", "3342179": "Pretty sure most of my answers correct :)\n\n**Pretrained weights / “publicly available”**\n\nI train almost all my models on local machines since I have 4 of them and almost no life :)\nI load them into datasets for the submission phase in a notebook.\nYou don't need to state anything right now - but if you are Gold and win some cash you will need to provide kaggle with evidence that you complied. Look in the Rules section - it describes. \nPretty much to win the cash prize you need to prove that you used only data that met the rules and that any pretrained weight models were available to the public. Outside of the money range there's no further proofs required.\n\n**Where training happens vs. 9h limit**\n\nNo big deal if training exceeds 9 but there is normally a limit on run time for any notebook - think its currently 12 hours.\nThis assumes that you train in one notebook and make the submission from another. If you working only on kaggle with no local training its pretty easy to link the saved models in the training notebook to the submission notebook. The submission notebook must meet the 9 hour rule. \n\n**Transparency expectations**\nThe Rules tab has a link to more of the requirements for winners with regards to documentation\n\nhttps://www.kaggle.com/WinningModelDocumentationGuidelines", "3342443": "I'm not a Kaggle official but I guess Jimmmy is correct.", "3343386": "HI @pcjimmmy , thanks a lot for your answer. So, as far as I understood , I can : \n\n- Use any public pre-trained model.\n- Fine-tune it however I want (local or Kaggle), as long as any extra data is public & allowed.\n- Store fine-tuned checkpoint in a private Kaggle dataset or notebook output.\n- Submission notebook: Attaches that dataset, Loads weights, Runs inference and writes submission.csv within 9h.\n- And more important, I only need to \"prove\" anything if in the prize range, and then the focus is on data sources + reproducibility, not on making tuned weights publicly visible to all competitors.\n\nN.B : At least you have 4 local machines to train your models, on my side, I only have access to kaggle GPU, and still no life 🥲" }, "source": "meta" }