{ "id": 663804, "title": "Some questions.", "url": "/competitions/deep-past-initiative-machine-translation/discussion/663804", "author_name": "", "post_date": "2025-12-20T13:50:43.893471100Z", "votes": 9, "comment_count": 7, "views": 0, "content": "

Hi @deeppast \nI have only a few questions. Any insights on them would be super helpful

\n

1) In the competition's overview page, can you please explain what this means ? we have included a lexicon in the dataset which includes all the proper nouns as specialists have normalized them for print publications

\n

2) How long can the hidden test set be? Any rough ball park number would be useful to plan the inference accordingly.

\n

3) For using open sourced models are there any specific whitelisting rules or all models found on the internet can be used? Is it the same one with 4000 samples in test.csv?

\n

4) From your end is it possible to add placeholder empty text in a new translation column in the test.csv? This to easily avoiding excess GPU hours being used while submission as the competition_rerun sometimes does not work for some reason.

", "messages": [ { "id": "3379718", "postDate": "12/20/2025 13:50:43", "content": "

Hi @deeppast \nI have only a few questions. Any insights on them would be super helpful

\n

1) In the competition's overview page, can you please explain what this means ? we have included a lexicon in the dataset which includes all the proper nouns as specialists have normalized them for print publications

\n

2) How long can the hidden test set be? Any rough ball park number would be useful to plan the inference accordingly.

\n

3) For using open sourced models are there any specific whitelisting rules or all models found on the internet can be used? Is it the same one with 4000 samples in test.csv?

\n

4) From your end is it possible to add placeholder empty text in a new translation column in the test.csv? This to easily avoiding excess GPU hours being used while submission as the competition_rerun sometimes does not work for some reason.

", "rawMarkdown": "Hi @deeppast \nI have only a few questions. Any insights on them would be super helpful\n\n1) In the competition's overview page, can you please explain what this means ? ``` we have included a lexicon in the dataset which includes all the proper nouns as specialists have normalized them for print publications ```\n\n2) How long can the hidden test set be? Any rough ball park number would be useful to plan the inference accordingly. \n\n3) For using open sourced models are there any specific whitelisting rules or all models found on the internet can be used? Is it the same one with 4000 samples in test.csv?\n\n4) From your end is it possible to add placeholder empty text in a new ```translation``` column in the test.csv? This to easily avoiding excess GPU hours being used while submission as the ```competition_rerun``` sometimes does not work for some reason.", "votes": null }, { "id": "3381024", "postDate": "12/23/2025 15:46:14", "content": "

Hi Aneesh,

\n

Thank you for your questions. I'm just discussing a couple of these points with the organizers.\nFor now, I can answer your first question:

\n

1) The lexicon includes proper nouns, i.e., personal names, place names, and institutions. These words are usually the most difficult for MT workflows, because they should not be translated as such, but only normalized. For example the name: Pu-šu-ki-in --> Pušu-ken is normalized as a personal name, rather than translated: \"his word is true\".

\n

For your other questions, I'll get back to you shortly!

", "rawMarkdown": "Hi Aneesh,\n\nThank you for your questions. I'm just discussing a couple of these points with the organizers.\nFor now, I can answer your first question:\n\n1) The lexicon includes proper nouns, i.e., personal names, place names, and institutions. These words are usually the most difficult for MT workflows, because they should not be translated as such, but only normalized. For example the name: Pu-šu-ki-in --> Pušu-ken is normalized as a personal name, rather than translated: \"his word is true\".\n\nFor your other questions, I'll get back to you shortly!", "votes": null }, { "id": "3381057", "postDate": "12/23/2025 16:48:17", "content": "

@deeppast
\nThank you for the response! \nI have few more doubts that I wanted to ask you without revealing or putting out my strategies publicly. May I reach you via email?

", "rawMarkdown": "deeppast \nThank you for the response! \nI have few more doubts that I wanted to ask you without revealing or putting out my strategies publicly. May I reach you via email?", "votes": null }, { "id": "3381125", "postDate": "12/23/2025 19:30:06", "content": "

How about we use Discord for DMs? https://discord.gg/XtCMyTrVCF

", "rawMarkdown": "How about we use Discord for DMs? https://discord.gg/XtCMyTrVCF", "votes": null }, { "id": "3381135", "postDate": "12/23/2025 20:08:33", "content": "

Hi @aneeshmukkamala,

\n

To answer your questions 2-4:

\n\n

Hope this answers your questions!

\n

Best,\nRyan

", "rawMarkdown": "Hi @aneeshmukkamala,\n\nTo answer your questions 2-4:\n- For the size of the test set, please refer to the competition's Data page, which states: \"There are about 4000 sentences in the test data from about 400 unique documents.\"\n- For the rules about allowed models and data, please see the competition's [Rules page](https://www.kaggle.com/competitions/deep-past-initiative-machine-translation/rules#5.-winner-license), sections 5 and subsequent.\n- I'm not sure I quite understand what you're asking for here, but in general my advice is to make your submission code robust to a variety of inputs and unexpected situations. Also, your personal GPU hours are not consumed by submission scoring, so you don't need to worry about submissions running down your GPU time.\n\nHope this answers your questions!\n\nBest,\nRyan", "votes": null }, { "id": "3381224", "postDate": "12/24/2025 03:54:11", "content": "

Thank you a lot for this @ryanholbrook \nRegarding question 4, I believe @deeppast can resolve this.\nWhat I meant is that for the 4 rows present in test.csv, we don't have the ground truth labels (translations). When we make the submission we can use them for debugging and comparing how our previous submissions went. It has been asked before , and due to the very limited number of data points even a couple of extra samples would make a difference

", "rawMarkdown": "Thank you a lot for this @ryanholbrook \nRegarding question 4, I believe @deeppast can resolve this.\nWhat I meant is that for the 4 rows present in test.csv, we don't have the ground truth labels (translations). When we make the submission we can use them for debugging and comparing how our previous submissions went. It has been asked [before ](https://www.kaggle.com/competitions/deep-past-initiative-machine-translation/discussion/663650), and due to the very limited number of data points even a couple of extra samples would make a difference", "votes": null }, { "id": "3382120", "postDate": "12/26/2025 15:13:51", "content": "

Hi @aneeshmukkamala, The sample sentences in test.csv were pulled from one of the translations in train.csv. I don't recall which one, but I imagine you could find it with a text search. We aren't able to share actual instances from the real test.csv unfortunately, but they did go through the same formatting and processing procedures as those in train.csv.

", "rawMarkdown": "Hi @aneeshmukkamala, The sample sentences in `test.csv` were pulled from one of the translations in `train.csv`. I don't recall which one, but I imagine you could find it with a text search. We aren't able to share actual instances from the real `test.csv` unfortunately, but they did go through the same formatting and processing procedures as those in `train.csv`.", "votes": null }, { "id": "3393324", "postDate": "01/18/2026 19:47:30", "content": "

@ryanholbrook and @deeppast, I can't find any of those four sentences (from test.csv) in train.csv. They also have the string i„ which I have do not find in train.csv.

", "rawMarkdown": "ryanholbrook and @deeppast, I can't find any of those four sentences (from `test.csv`) in `train.csv`. They also have the string `i„` which I have do not find in `train.csv`.", "votes": null } ], "comments": [ { "id": 3381024, "author_name": "deeppast", "author_url": "", "post_date": "12/23/2025 15:46:14", "content": "

Hi Aneesh,

\n

Thank you for your questions. I'm just discussing a couple of these points with the organizers.\nFor now, I can answer your first question:

\n

1) The lexicon includes proper nouns, i.e., personal names, place names, and institutions. These words are usually the most difficult for MT workflows, because they should not be translated as such, but only normalized. For example the name: Pu-šu-ki-in --> Pušu-ken is normalized as a personal name, rather than translated: \"his word is true\".

\n

For your other questions, I'll get back to you shortly!

", "votes": null, "replies": [ { "id": 3381057, "author_name": "", "author_url": "", "post_date": "12/23/2025 16:48:17", "content": "

@deeppast
\nThank you for the response! \nI have few more doubts that I wanted to ask you without revealing or putting out my strategies publicly. May I reach you via email?

", "votes": null, "replies": [ { "id": 3381125, "author_name": "deeppast", "author_url": "", "post_date": "12/23/2025 19:30:06", "content": "

How about we use Discord for DMs? https://discord.gg/XtCMyTrVCF

", "votes": null, "replies": [] } ] } ] }, { "id": 3381135, "author_name": "ryanholbrook", "author_url": "", "post_date": "12/23/2025 20:08:33", "content": "

Hi @aneeshmukkamala,

\n

To answer your questions 2-4:

\n\n

Hope this answers your questions!

\n

Best,\nRyan

", "votes": null, "replies": [ { "id": 3381224, "author_name": "", "author_url": "", "post_date": "12/24/2025 03:54:11", "content": "

Thank you a lot for this @ryanholbrook \nRegarding question 4, I believe @deeppast can resolve this.\nWhat I meant is that for the 4 rows present in test.csv, we don't have the ground truth labels (translations). When we make the submission we can use them for debugging and comparing how our previous submissions went. It has been asked before , and due to the very limited number of data points even a couple of extra samples would make a difference

", "votes": null, "replies": [ { "id": 3382120, "author_name": "ryanholbrook", "author_url": "", "post_date": "12/26/2025 15:13:51", "content": "

Hi @aneeshmukkamala, The sample sentences in test.csv were pulled from one of the translations in train.csv. I don't recall which one, but I imagine you could find it with a text search. We aren't able to share actual instances from the real test.csv unfortunately, but they did go through the same formatting and processing procedures as those in train.csv.

", "votes": null, "replies": [ { "id": 3393324, "author_name": "einherjer", "author_url": "", "post_date": "01/18/2026 19:47:30", "content": "

@ryanholbrook and @deeppast, I can't find any of those four sentences (from test.csv) in train.csv. They also have the string i„ which I have do not find in train.csv.

", "votes": null, "replies": [] } ] } ] } ] } ], "raw_markdown_by_id": { "3379718": "Hi @deeppast \nI have only a few questions. Any insights on them would be super helpful\n\n1) In the competition's overview page, can you please explain what this means ? ``` we have included a lexicon in the dataset which includes all the proper nouns as specialists have normalized them for print publications ```\n\n2) How long can the hidden test set be? Any rough ball park number would be useful to plan the inference accordingly. \n\n3) For using open sourced models are there any specific whitelisting rules or all models found on the internet can be used? Is it the same one with 4000 samples in test.csv?\n\n4) From your end is it possible to add placeholder empty text in a new ```translation``` column in the test.csv? This to easily avoiding excess GPU hours being used while submission as the ```competition_rerun``` sometimes does not work for some reason.", "3381024": "Hi Aneesh,\n\nThank you for your questions. I'm just discussing a couple of these points with the organizers.\nFor now, I can answer your first question:\n\n1) The lexicon includes proper nouns, i.e., personal names, place names, and institutions. These words are usually the most difficult for MT workflows, because they should not be translated as such, but only normalized. For example the name: Pu-šu-ki-in --> Pušu-ken is normalized as a personal name, rather than translated: \"his word is true\".\n\nFor your other questions, I'll get back to you shortly!", "3381057": "deeppast \nThank you for the response! \nI have few more doubts that I wanted to ask you without revealing or putting out my strategies publicly. May I reach you via email?", "3381125": "How about we use Discord for DMs? https://discord.gg/XtCMyTrVCF", "3381135": "Hi @aneeshmukkamala,\n\nTo answer your questions 2-4:\n- For the size of the test set, please refer to the competition's Data page, which states: \"There are about 4000 sentences in the test data from about 400 unique documents.\"\n- For the rules about allowed models and data, please see the competition's [Rules page](https://www.kaggle.com/competitions/deep-past-initiative-machine-translation/rules#5.-winner-license), sections 5 and subsequent.\n- I'm not sure I quite understand what you're asking for here, but in general my advice is to make your submission code robust to a variety of inputs and unexpected situations. Also, your personal GPU hours are not consumed by submission scoring, so you don't need to worry about submissions running down your GPU time.\n\nHope this answers your questions!\n\nBest,\nRyan", "3381224": "Thank you a lot for this @ryanholbrook \nRegarding question 4, I believe @deeppast can resolve this.\nWhat I meant is that for the 4 rows present in test.csv, we don't have the ground truth labels (translations). When we make the submission we can use them for debugging and comparing how our previous submissions went. It has been asked [before ](https://www.kaggle.com/competitions/deep-past-initiative-machine-translation/discussion/663650), and due to the very limited number of data points even a couple of extra samples would make a difference", "3382120": "Hi @aneeshmukkamala, The sample sentences in `test.csv` were pulled from one of the translations in `train.csv`. I don't recall which one, but I imagine you could find it with a text search. We aren't able to share actual instances from the real `test.csv` unfortunately, but they did go through the same formatting and processing procedures as those in `train.csv`.", "3393324": "ryanholbrook and @deeppast, I can't find any of those four sentences (from `test.csv`) in `train.csv`. They also have the string `i„` which I have do not find in `train.csv`." }, "source": "meta" }