If you feel stuck or have any questions, here are all the discussions you should read over:
\nYou should also review the dataset instructions section of the overview page.
", "messages": [ { "id": 3392252, "postDate": "2026-01-16T15:04:51.413Z", "content": "If you feel stuck or have any questions, here are all the discussions you should read over:
\nYou should also review the dataset instructions section of the overview page.
", "rawMarkdown": "If you feel stuck or have any questions, here are all the discussions you should read over:\n- [Two practical stumbling blocks in Akkadian → English MT (and how to address them)](https://www.kaggle.com/competitions/deep-past-initiative-machine-translation/discussion/665209)\n- [The translation column of the final test predictions into the correspondingDiscussion on the Old Assyrian calendar and how to handle month names
", "rawMarkdown": "[Discussion on the Old Assyrian calendar and how to handle month names](https://www.kaggle.com/competitions/deep-past-initiative-machine-translation/discussion/667480#3390747)", "votes": 1 }, { "id": 3392860, "postDate": "2026-01-17T16:52:16.440Z", "content": "Hi @jackvd,
\nThank you for compiling and sharing these discussions 🙏. I’ve read through them carefully and they were genuinely helpful, especially in clarifying which approaches tend to hurt performance rather than help.
\nI wanted to ask a few quick clarification questions, if you don’t mind. I’ve trained several models now and while training loss keeps improving, my leaderboard score seems to have plateaued, so I’m trying to understand where the real leverage is.
\nTraining data: Did you train using only train.csv, or we have to incorporate any external sources (e.g. ORACC / MTM24 / published_texts)?
\nSentence alignment: Did you perform sentence-level alignment on all ~1,581 documents, or only on the longer documents that exceed the 512 token limit?
\nḫ / h handling: In the training data, did you normalize ḫ -> h (and Ḫ -> H) before training, or keep the original forms and rely on the model?
\nModel strategy: Did you focus on a single strong model, or did you find ensembling multiple models to be necessary? I can ensemble my best checkpoints, but I’m wondering whether a single well-trained model is already sufficient.
\nThanks again for taking the time to share your insights. They’ve helped cut through a lot of confusing or misleading advice, and I really appreciate it.
", "rawMarkdown": "**Hi** @jackvd,\n\nThank you for compiling and sharing these discussions 🙏. I’ve read through them carefully and they were genuinely helpful, especially in clarifying which approaches tend to hurt performance rather than help.\n\nI wanted to ask a few quick clarification questions, if you don’t mind. I’ve trained several models now and while training loss keeps improving, my leaderboard score seems to have plateaued, so I’m trying to understand where the real leverage is.\n\n**Training data**: Did you train using only **train.csv**, or we have to incorporate any external sources (e.g. ORACC / MTM24 / published_texts)?\n\n**Sentence alignment**: Did you perform sentence-level alignment on all ~1,581 documents, or only on the longer documents that exceed the 512 token limit?\n\n**ḫ / h handling**: In the training data, did you normalize **ḫ** -> **h** (and **Ḫ -> H**) before training, or keep the original forms and rely on the model?\n\n**Model strategy**: Did you focus on a single strong model, or did you find ensembling multiple models to be necessary? I can ensemble my best checkpoints, but I’m wondering whether a single well-trained model is already sufficient.\n\nThanks again for taking the time to share your insights. They’ve helped cut through a lot of confusing or misleading advice, and I really appreciate it.", "votes": 1 }, { "id": 3392295, "postDate": "2026-01-16T16:36:29.220Z", "content": "Thanks JACK! These are all excellent guides for those wishing to increase their %
\nIf we consider that the document level data needs to be processed in terms of sentences in alignment, that will probably be what separates the top leaders from the rest.
", "rawMarkdown": "Thanks JACK! These are all excellent guides for those wishing to increase their %\n\nIf we consider that the document level data needs to be processed in terms of sentences in alignment, that will probably be what separates the top leaders from the rest.", "votes": 1, "replies": [ { "id": 3392349, "postDate": "2026-01-16T18:38:12.150Z", "content": "Hi @deeppast ,
\nDo you have an answer for dot h/H?\nhttps://www.kaggle.com/competitions/deep-past-initiative-machine-translation/discussion/665949
\nOr maybe you cannot disclose such information on how you preprocess dot h/H?
\nCurrently I'm dealing with it as augmentation, I keep dot h/H as is and the I replace randomly by regular h/H with an augmentation process. But I'm quite sure it could be better.
\nThanks.
", "rawMarkdown": "Hi @deeppast ,\n\nDo you have an answer for dot h/H?\nhttps://www.kaggle.com/competitions/deep-past-initiative-machine-translation/discussion/665949\n\nOr maybe you cannot disclose such information on how you preprocess dot h/H?\n\nCurrently I'm dealing with it as augmentation, I keep dot h/H as is and the I replace randomly by regular h/H with an augmentation process. But I'm quite sure it could be better.\n\nThanks.", "replies": [ { "id": 3392350, "postDate": "2026-01-16T18:38:40.217Z", "content": "Replace them with h/H.
", "rawMarkdown": "Replace them with h/H.", "votes": 3 }, { "id": 3392504, "postDate": "2026-01-17T04:26:43.740Z", "content": "Right, there are no dot h/H or hooked h/H. There's only regular h/H in the test.
\nYou can see that's the case in the character set I posted: https://www.kaggle.com/competitions/deep-past-initiative-machine-translation/discussion/665209#3384200
", "rawMarkdown": "Right, there are no dot h/H or hooked h/H. There's only regular h/H in the test. \n\nYou can see that's the case in the character set I posted: https://www.kaggle.com/competitions/deep-past-initiative-machine-translation/discussion/665209#3384200", "votes": 1 }, { "id": 3392623, "postDate": "2026-01-17T08:49:11.087Z", "content": "\n\nRight, there are no dot h/H or hooked h/H. There's only regular h/H in the test.
\nYou can see that's the case in the character set I posted: https://www.kaggle.com/competitions/deep-past-initiative-machine-translation/discussion/665209#3384200
\n
We should probably get the h/H corrected in the Overview character table. Also, I'm a bit suspicious of the complete lack of subscripted X (ₓ) in the entire test set. Since all numerical subscripts are turned into digits, but X should be retained then either:
<gap>.Could we rule out the second possibility by verifying the data transformation pipeline?
", "rawMarkdown": "> Right, there are no dot h/H or hooked h/H. There's only regular h/H in the test. \n> \n> You can see that's the case in the character set I posted: https://www.kaggle.com/competitions/deep-past-initiative-machine-translation/discussion/665209#3384200\n\nWe should probably get the h/H corrected in the Overview character table. Also, I'm a bit suspicious of the complete lack of subscripted X (`ₓ`) in the entire test set. Since all numerical subscripts are turned into digits, but X should be retained then either:\n- there is not a single subscripted X in the entire test set\n- the subscripted X was turned into a regular sized X and transformed into `Finally I got it working after several attempts, single model +0.4 on LB by training with h/H.
", "rawMarkdown": "Finally I got it working after several attempts, single model +0.4 on LB by training with h/H. " }, { "id": 3394505, "postDate": "2026-01-21T06:35:48.357Z", "content": "Is there something other than just converting ḫ/Ḫ -> h/H on both transliteration and translation.
", "rawMarkdown": "Is there something other than just converting ḫ/Ḫ -> h/H on both transliteration and translation. " }, { "id": 3394517, "postDate": "2026-01-21T07:28:24.550Z", "content": "@mpware is it encoding related? I simply replaced h/H with underlines to regular h/H in my train data few days ago and got -1.4 LB
", "rawMarkdown": "@mpware is it encoding related? I simply replaced h/H with underlines to regular h/H in my train data few days ago and got -1.4 LB" }, { "id": 3394527, "postDate": "2026-01-21T07:53:50.453Z", "content": "Yes, just replaced:
\nTRANSLIT_SPECIAL_CHAR_MAP = {\n # h-dot does not exist in test set\n \"ḫ\": \"h\",\n \"Ḫ\": \"H\",\n \"ʾ\": \"\",\n # Same for these\n \"mₓ\": \"m\",\n \"zₓ\": \"z\",\n ...\n}\nI'm also using some random augmentations (low probability) to replace some diacritics to simulate noise:
\nDIACRITIC_COMMON_MAP = {\n \"š\": \"s\", \"Š\": \"S\",\n...\nI had to test a few checkpoints before getting +0.4.
", "rawMarkdown": "Yes, just replaced:\n\n```python\nTRANSLIT_SPECIAL_CHAR_MAP = {\n # h-dot does not exist in test set\n \"ḫ\": \"h\",\n \"Ḫ\": \"H\",\n \"ʾ\": \"\",\n # Same for these\n \"mₓ\": \"m\",\n \"zₓ\": \"z\",\n ...\n}\n```\n\nI'm also using some random augmentations (low probability) to replace some diacritics to simulate noise:\n\n```python\nDIACRITIC_COMMON_MAP = {\n \"š\": \"s\", \"Š\": \"S\",\n...\n```\n\nI had to test a few checkpoints before getting +0.4.", "votes": 6 }, { "id": 3394539, "postDate": "2026-01-21T08:25:29.697Z", "content": "Great, I started out replacing ḫ/Ḫ -> h/H and didn't experiment without them, I was just some what confident with the replacement \nI haven't tried random augmenting till now , will check that helps improve the LB score or not
\nAlso I can't clearly map out a relation between my CV and LB , model trained on sentences score more than the model trained on docs on my CV (4-5 point difference) , however there is a max to max difference of +/- 0.1 on LB which is mostly variance
\nAlso my CV is a subset of sentence aligned which I got from splitting from the sen_oare csv file.
\nThank you , I'm learning a lot
", "rawMarkdown": "Great, I started out replacing ḫ/Ḫ -> h/H and didn't experiment without them, I was just some what confident with the replacement \nI haven't tried random augmenting till now , will check that helps improve the LB score or not \n\nAlso I can't clearly map out a relation between my CV and LB , model trained on sentences score more than the model trained on docs on my CV (4-5 point difference) , however there is a max to max difference of +/- 0.1 on LB which is mostly variance \n\nAlso my CV is a subset of sentence aligned which I got from splitting from the sen_oare csv file. \n\nThank you , I'm learning a lot ", "votes": 3 }, { "id": 3394542, "postDate": "2026-01-21T08:39:35.883Z", "content": "Oh alr. Thanks for sharing!
", "rawMarkdown": "Oh alr. Thanks for sharing!" } ] }, { "id": 3392858, "postDate": "2026-01-17T16:47:13.667Z", "rawMarkdown": "", "isDeleted": true } ] }, { "id": 3395878, "postDate": "2026-01-23T20:49:15.840Z", "content": "HEY SIR\ntokenizer(\"a-na i-tù-ra-ma\")\nIm running this and its taking too much time. what could be the problem?
", "rawMarkdown": "HEY SIR\ntokenizer(\"a-na i-tù-ra-ma\")\nIm running this and its taking too much time. what could be the problem?", "votes": -2 }, { "id": 3409233, "postDate": "2026-02-22T12:18:46.397Z", "content": "Thank you for sharing and curating this information
", "rawMarkdown": "Thank you for sharing and curating this information " }, { "id": 3406970, "postDate": "2026-02-17T08:32:17.060Z", "content": "@deeppast @ryanholbrook \nThis thread is a good source of typical stumbling blocks with answers and insights on their mitigation. Could we get it pinned along with the \"Typical stumbling blocks\" thread? It could help spreading the base knowledge the Community has come up with.
", "rawMarkdown": "@deeppast @ryanholbrook \nThis thread is a good source of typical stumbling blocks with answers and insights on their mitigation. Could we get it pinned along with the \"Typical stumbling blocks\" thread? It could help spreading the base knowledge the Community has come up with." }, { "id": 3392597, "postDate": "2026-01-17T07:29:46.357Z", "content": "Avoid Bad Advice? @jackvd
", "rawMarkdown": "Avoid Bad Advice? @jackvd ", "replies": [ { "id": 3392828, "postDate": "2026-01-17T15:32:33.003Z", "content": "I’ve seen some notebooks and discussions giving bad advice to new competitors or people struggling. I’m hoping this helps avoid that problem and get more people on the right path to improving
", "rawMarkdown": "I’ve seen some notebooks and discussions giving bad advice to new competitors or people struggling. I’m hoping this helps avoid that problem and get more people on the right path to improving", "votes": 6, "replies": [ { "id": 3392852, "postDate": "2026-01-17T16:28:59.453Z", "rawMarkdown": "", "isDeleted": true }, { "id": 3394704, "postDate": "2026-01-21T14:41:25.540Z", "content": "Should we be using pre-trained models, or should we be fine-tuning models like byT5 ourselves?
", "rawMarkdown": "Should we be using pre-trained models, or should we be fine-tuning models like byT5 ourselves?" } ] } ] }, { "id": 3392343, "postDate": "2026-01-16T18:30:18.627Z", "content": "@jackvd \nConsidering the Incomplete translations topic this discussion gives more potential for data cleaning. It also sheds light on the problem of incomplte transliterations.
", "rawMarkdown": "@jackvd \nConsidering the **Incomplete translations** topic [this discussion](https://www.kaggle.com/competitions/deep-past-initiative-machine-translation/discussion/667483) gives more potential for data cleaning. It also sheds light on the problem of incomplte transliterations." } ], "comments": [ { "id": 3394526, "author_name": "Angantyr", "author_url": "", "post_date": "2026-01-21T07:53:10.560000", "content": "Discussion on the Old Assyrian calendar and how to handle month names
", "votes": 1, "replies": [] }, { "id": 3392860, "author_name": "Assia Benkedia", "author_url": "", "post_date": "2026-01-17T16:52:16.440000", "content": "Hi @jackvd,
\nThank you for compiling and sharing these discussions 🙏. I’ve read through them carefully and they were genuinely helpful, especially in clarifying which approaches tend to hurt performance rather than help.
\nI wanted to ask a few quick clarification questions, if you don’t mind. I’ve trained several models now and while training loss keeps improving, my leaderboard score seems to have plateaued, so I’m trying to understand where the real leverage is.
\nTraining data: Did you train using only train.csv, or we have to incorporate any external sources (e.g. ORACC / MTM24 / published_texts)?
\nSentence alignment: Did you perform sentence-level alignment on all ~1,581 documents, or only on the longer documents that exceed the 512 token limit?
\nḫ / h handling: In the training data, did you normalize ḫ -> h (and Ḫ -> H) before training, or keep the original forms and rely on the model?
\nModel strategy: Did you focus on a single strong model, or did you find ensembling multiple models to be necessary? I can ensemble my best checkpoints, but I’m wondering whether a single well-trained model is already sufficient.
\nThanks again for taking the time to share your insights. They’ve helped cut through a lot of confusing or misleading advice, and I really appreciate it.
", "votes": 1, "replies": [] }, { "id": 3392295, "author_name": "Adam Anderson", "author_url": "", "post_date": "2026-01-16T16:36:29.220000", "content": "Thanks JACK! These are all excellent guides for those wishing to increase their %
\nIf we consider that the document level data needs to be processed in terms of sentences in alignment, that will probably be what separates the top leaders from the rest.
", "votes": 1, "replies": [ { "id": 3392349, "author_name": "MPWARE", "author_url": "", "post_date": "2026-01-16T18:38:12.150000", "content": "Hi @deeppast ,
\nDo you have an answer for dot h/H?\nhttps://www.kaggle.com/competitions/deep-past-initiative-machine-translation/discussion/665949
\nOr maybe you cannot disclose such information on how you preprocess dot h/H?
\nCurrently I'm dealing with it as augmentation, I keep dot h/H as is and the I replace randomly by regular h/H with an augmentation process. But I'm quite sure it could be better.
\nThanks.
", "votes": 0, "replies": [ { "id": 3392350, "author_name": "Jack", "author_url": "", "post_date": "2026-01-16T18:38:40.217000", "content": "Replace them with h/H.
", "votes": 3, "replies": [] }, { "id": 3392504, "author_name": "Adam Anderson", "author_url": "", "post_date": "2026-01-17T04:26:43.740000", "content": "Right, there are no dot h/H or hooked h/H. There's only regular h/H in the test.
\nYou can see that's the case in the character set I posted: https://www.kaggle.com/competitions/deep-past-initiative-machine-translation/discussion/665209#3384200
", "votes": 1, "replies": [] }, { "id": 3392623, "author_name": "Angantyr", "author_url": "", "post_date": "2026-01-17T08:49:11.087000", "content": "\n\nRight, there are no dot h/H or hooked h/H. There's only regular h/H in the test.
\nYou can see that's the case in the character set I posted: https://www.kaggle.com/competitions/deep-past-initiative-machine-translation/discussion/665209#3384200
\n
We should probably get the h/H corrected in the Overview character table. Also, I'm a bit suspicious of the complete lack of subscripted X (ₓ) in the entire test set. Since all numerical subscripts are turned into digits, but X should be retained then either:
<gap>.Could we rule out the second possibility by verifying the data transformation pipeline?
", "votes": 3, "replies": [] }, { "id": 3394332, "author_name": "MPWARE", "author_url": "", "post_date": "2026-01-20T21:13:44.403000", "content": "Finally I got it working after several attempts, single model +0.4 on LB by training with h/H.
", "votes": 0, "replies": [] }, { "id": 3394505, "author_name": "Hrithik Reddy", "author_url": "", "post_date": "2026-01-21T06:35:48.357000", "content": "Is there something other than just converting ḫ/Ḫ -> h/H on both transliteration and translation.
", "votes": 0, "replies": [] }, { "id": 3394517, "author_name": "Yurnero", "author_url": "", "post_date": "2026-01-21T07:28:24.550000", "content": "@mpware is it encoding related? I simply replaced h/H with underlines to regular h/H in my train data few days ago and got -1.4 LB
", "votes": 0, "replies": [] }, { "id": 3394527, "author_name": "MPWARE", "author_url": "", "post_date": "2026-01-21T07:53:50.453000", "content": "Yes, just replaced:
\nTRANSLIT_SPECIAL_CHAR_MAP = {\n # h-dot does not exist in test set\n \"ḫ\": \"h\",\n \"Ḫ\": \"H\",\n \"ʾ\": \"\",\n # Same for these\n \"mₓ\": \"m\",\n \"zₓ\": \"z\",\n ...\n}\nI'm also using some random augmentations (low probability) to replace some diacritics to simulate noise:
\nDIACRITIC_COMMON_MAP = {\n \"š\": \"s\", \"Š\": \"S\",\n...\nI had to test a few checkpoints before getting +0.4.
", "votes": 6, "replies": [] }, { "id": 3394539, "author_name": "Hrithik Reddy", "author_url": "", "post_date": "2026-01-21T08:25:29.697000", "content": "Great, I started out replacing ḫ/Ḫ -> h/H and didn't experiment without them, I was just some what confident with the replacement \nI haven't tried random augmenting till now , will check that helps improve the LB score or not
\nAlso I can't clearly map out a relation between my CV and LB , model trained on sentences score more than the model trained on docs on my CV (4-5 point difference) , however there is a max to max difference of +/- 0.1 on LB which is mostly variance
\nAlso my CV is a subset of sentence aligned which I got from splitting from the sen_oare csv file.
\nThank you , I'm learning a lot
", "votes": 3, "replies": [] }, { "id": 3394542, "author_name": "Yurnero", "author_url": "", "post_date": "2026-01-21T08:39:35.883000", "content": "Oh alr. Thanks for sharing!
", "votes": 0, "replies": [] } ] }, { "id": 3392858, "author_name": "", "author_url": "", "post_date": "2026-01-17T16:47:13.667000", "content": "", "votes": 0, "replies": [] } ] }, { "id": 3395878, "author_name": "Simo Medamro", "author_url": "", "post_date": "2026-01-23T20:49:15.840000", "content": "HEY SIR\ntokenizer(\"a-na i-tù-ra-ma\")\nIm running this and its taking too much time. what could be the problem?
", "votes": -2, "replies": [] }, { "id": 3409233, "author_name": "Rami Ismael", "author_url": "", "post_date": "2026-02-22T12:18:46.397000", "content": "Thank you for sharing and curating this information
", "votes": 0, "replies": [] }, { "id": 3406970, "author_name": "Angantyr", "author_url": "", "post_date": "2026-02-17T08:32:17.060000", "content": "@deeppast @ryanholbrook \nThis thread is a good source of typical stumbling blocks with answers and insights on their mitigation. Could we get it pinned along with the \"Typical stumbling blocks\" thread? It could help spreading the base knowledge the Community has come up with.
", "votes": 0, "replies": [] }, { "id": 3392597, "author_name": "Navneet", "author_url": "", "post_date": "2026-01-17T07:29:46.357000", "content": "Avoid Bad Advice? @jackvd
", "votes": 0, "replies": [ { "id": 3392828, "author_name": "Jack", "author_url": "", "post_date": "2026-01-17T15:32:33.003000", "content": "I’ve seen some notebooks and discussions giving bad advice to new competitors or people struggling. I’m hoping this helps avoid that problem and get more people on the right path to improving
", "votes": 6, "replies": [ { "id": 3392852, "author_name": "", "author_url": "", "post_date": "2026-01-17T16:28:59.453000", "content": "", "votes": 0, "replies": [] }, { "id": 3394704, "author_name": "Leo Withrow", "author_url": "", "post_date": "2026-01-21T14:41:25.540000", "content": "Should we be using pre-trained models, or should we be fine-tuning models like byT5 ourselves?
", "votes": 0, "replies": [] } ] } ] }, { "id": 3392343, "author_name": "Angantyr", "author_url": "", "post_date": "2026-01-16T18:30:18.627000", "content": "@jackvd \nConsidering the Incomplete translations topic this discussion gives more potential for data cleaning. It also sheds light on the problem of incomplte transliterations.
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "3392252": "If you feel stuck or have any questions, here are all the discussions you should read over:\n- [Two practical stumbling blocks in Akkadian → English MT (and how to address them)](https://www.kaggle.com/competitions/deep-past-initiative-machine-translation/discussion/665209)\n- [The translation column of the final test predictions into the corresponding