First and foremost, I would like to express my gratitude to the organizers for hosting this excellent Akkadian translation competition (even though we encountered—and subsequently resolved—various challenges throughout the event). Despite these hurdles, it remained a truly remarkable contest !
\nI would like to give special thanks to one of my teammates, @colhomy , who invested a tremendous amount of effort into training the models and cleaning the data during the competition. I also wish to thank my other teammates for the diverse ideas and approaches they contributed during the early stages of the competition, which were instrumental in helping us achieve such a successful result. \nAt the end of, I would like to thank @giovannyrodrguez 's code. his code provides a robust ensemble method.
\n\"translate Akkadian to English: \")<gap> and decimal processing in the preprocessing stage, whereas some public code versions put these steps in post-processing.⅚ form, I also handled the 5/6 forms in the data, processing both decimal and 5/6 formats into the ⅚ form.-textiles —> kutānum textiles, etc.).Gemini-3.0-flash-preview was used for splitting, correction, pseudo-labeling, and PDF identification.
\nUsed Gemini-3.0-flash-preview to split train.csv to avoid the 512-token input limit for the prefix + transliteration.
action label for each entry:keep: Kept in original form, no deletions made.trim: Deleted a part (trimmed redundant or non-corresponding content).delete: Entire entry invalid or unable to align, suggested for discard (source and translation do not correspond at all).OA_Lexicon_eBL.csv, onomasticon.csv, and train.csv to extract transliteration-to-replacement pairs.train.csv; if a transliteration had multiple translation variants, only the most frequent one appearing in train.csv was used.publication.csv, the same proper noun correction method was applied. Gemini was used to extract paragraphs containing both source and translation.Due to model instability, I performed a final cleanup of erroneous samples, including empty values and samples where the source was more than twice as long as the translation (source is usually shorter than translation).
\ntrain.csv was effective for scores below 35. However, after exceeding 36, fine-tuning on train.csv caused a drop in LB score. I switched to training on the mixed dataset directly.group_by_length enabled.num_beam_cands = 4, num_beams = 8.First and foremost, I would like to express my gratitude to the organizers for hosting this excellent Akkadian translation competition (even though we encountered—and subsequently resolved—various challenges throughout the event). Despite these hurdles, it remained a truly remarkable contest !
\nI would like to give special thanks to one of my teammates, @colhomy , who invested a tremendous amount of effort into training the models and cleaning the data during the competition. I also wish to thank my other teammates for the diverse ideas and approaches they contributed during the early stages of the competition, which were instrumental in helping us achieve such a successful result. \nAt the end of, I would like to thank @giovannyrodrguez 's code. his code provides a robust ensemble method.
\n\"translate Akkadian to English: \")<gap> and decimal processing in the preprocessing stage, whereas some public code versions put these steps in post-processing.⅚ form, I also handled the 5/6 forms in the data, processing both decimal and 5/6 formats into the ⅚ form.-textiles —> kutānum textiles, etc.).Gemini-3.0-flash-preview was used for splitting, correction, pseudo-labeling, and PDF identification.
\nUsed Gemini-3.0-flash-preview to split train.csv to avoid the 512-token input limit for the prefix + transliteration.
action label for each entry:keep: Kept in original form, no deletions made.trim: Deleted a part (trimmed redundant or non-corresponding content).delete: Entire entry invalid or unable to align, suggested for discard (source and translation do not correspond at all).OA_Lexicon_eBL.csv, onomasticon.csv, and train.csv to extract transliteration-to-replacement pairs.train.csv; if a transliteration had multiple translation variants, only the most frequent one appearing in train.csv was used.publication.csv, the same proper noun correction method was applied. Gemini was used to extract paragraphs containing both source and translation.Due to model instability, I performed a final cleanup of erroneous samples, including empty values and samples where the source was more than twice as long as the translation (source is usually shorter than translation).
\ntrain.csv was effective for scores below 35. However, after exceeding 36, fine-tuning on train.csv caused a drop in LB score. I switched to training on the mixed dataset directly.group_by_length enabled.num_beam_cands = 4, num_beams = 8.great, wonderful approach , can you share your training code
", "rawMarkdown": "great, wonderful approach , can you share your training code", "votes": null }, { "id": "3427582", "postDate": "03/24/2026 09:11:10", "content": "My training code is similar to the public DPC Starter Train; the differences are minimal, as the primary focus was on the data.
", "rawMarkdown": "My training code is similar to the public DPC Starter Train; the differences are minimal, as the primary focus was on the data.", "votes": null } ], "comments": [ { "id": 3427538, "author_name": "nacefmbarek", "author_url": "", "post_date": "03/24/2026 08:02:06", "content": "great, wonderful approach , can you share your training code
", "votes": null, "replies": [ { "id": 3427582, "author_name": "colhomy", "author_url": "", "post_date": "03/24/2026 09:11:10", "content": "My training code is similar to the public DPC Starter Train; the differences are minimal, as the primary focus was on the data.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "3427406": "## 0.Preface\nFirst and foremost, I would like to express my gratitude to the organizers for hosting this excellent Akkadian translation competition (even though we encountered—and subsequently resolved—various challenges throughout the event). Despite these hurdles, it remained a truly remarkable contest !\n\n I would like to give special thanks to one of my teammates, @colhomy , who invested a tremendous amount of effort into training the models and cleaning the data during the competition. I also wish to thank my other teammates for the diverse ideas and approaches they contributed during the early stages of the competition, which were instrumental in helping us achieve such a successful result. \nAt the end of, I would like to thank @giovannyrodrguez 's [code](https://www.kaggle.com/code/giovannyrodrguez/lb-35-9-ensembling-post-processing-baseline). his code provides a robust ensemble method.\n\n## 1. Overview\n- **Model**: ByT5-base only (Prefix: `\"translate Akkadian to English: \"`)\n- **Hardware**: Single RTX 4090\n- **Key Strategy**: For me, the only reliable way to consistently improve the score in this competition was increasing the data volume.\n\n## 2. Pre-processing & Post-processing\n- **Handling**: Processed based on discussion information, similar to public code but with key differences.\n- **Preprocessing Details**: I used full `