This was a challenging competition, seemingly made more difficult than necessary by the mis-alignment of the train and test sets. Nevertheless, it was quite fascinating and I became a bit obsessed with it (as my wife will attest!).
\nI spent a lot of time on data preprocessing, extraction, and sentence splitting. But the most challenging part for me was searching for a validation set that I could rely on. I eventually ended up with a group for tablets from the train set, AKT 8, AKT 12, AKT 6e, and two recent books by Cécile Michel, all restricted to sentence-level pairs. Of these, the tablets that seemed to be closest to the test set were the texts in Michel's \"Women of Aššur and Kaneš - Texts from the Archives of Assyrian Merchants\" and, ironically, the ones furthest from the test set were the texts from train.csv. But there were still significant gaps between my valid and test scores. My best model scored 42.1 valid, 36.3 public LB, 38.1 private LB.
\nOne technique I used, which I haven't seen mentioned yet, was to include each sentence's starting line number as well as surrounding text (before and after each sentence ) as additional context in the prompt.
\nLooking at the solutions published so far, I can see I missed a number of opportunities to improve my solution. Congratulations to the winners!
", "messages": [ { "id": 3427581, "postDate": "2026-03-24T09:10:00.547Z", "content": "This was a challenging competition, seemingly made more difficult than necessary by the mis-alignment of the train and test sets. Nevertheless, it was quite fascinating and I became a bit obsessed with it (as my wife will attest!).
\nI spent a lot of time on data preprocessing, extraction, and sentence splitting. But the most challenging part for me was searching for a validation set that I could rely on. I eventually ended up with a group for tablets from the train set, AKT 8, AKT 12, AKT 6e, and two recent books by Cécile Michel, all restricted to sentence-level pairs. Of these, the tablets that seemed to be closest to the test set were the texts in Michel's \"Women of Aššur and Kaneš - Texts from the Archives of Assyrian Merchants\" and, ironically, the ones furthest from the test set were the texts from train.csv. But there were still significant gaps between my valid and test scores. My best model scored 42.1 valid, 36.3 public LB, 38.1 private LB.
\nOne technique I used, which I haven't seen mentioned yet, was to include each sentence's starting line number as well as surrounding text (before and after each sentence ) as additional context in the prompt.
\nLooking at the solutions published so far, I can see I missed a number of opportunities to improve my solution. Congratulations to the winners!
", "rawMarkdown": "This was a challenging competition, seemingly made more difficult than necessary by the mis-alignment of the train and test sets. Nevertheless, it was quite fascinating and I became a bit obsessed with it (as my wife will attest!). \n\nI spent a lot of time on data preprocessing, extraction, and sentence splitting. But the most challenging part for me was searching for a validation set that I could rely on. I eventually ended up with a group for tablets from the train set, AKT 8, AKT 12, AKT 6e, and two recent books by Cécile Michel, all restricted to sentence-level pairs. Of these, the tablets that seemed to be closest to the test set were the texts in Michel's \"Women of Aššur and Kaneš - Texts from the Archives of Assyrian Merchants\" and, ironically, the ones furthest from the test set were the texts from train.csv. But there were still significant gaps between my valid and test scores. My best model scored 42.1 valid, 36.3 public LB, 38.1 private LB.\n\nOne technique I used, which I haven't seen mentioned yet, was to include each sentence's starting line number as well as surrounding text (before and after each sentence ) as additional context in the prompt. \n\nLooking at the solutions published so far, I can see I missed a number of opportunities to improve my solution. Congratulations to the winners!", "votes": 10 } ], "comments": [], "raw_markdown_by_id": { "3427581": "This was a challenging competition, seemingly made more difficult than necessary by the mis-alignment of the train and test sets. Nevertheless, it was quite fascinating and I became a bit obsessed with it (as my wife will attest!). \n\nI spent a lot of time on data preprocessing, extraction, and sentence splitting. But the most challenging part for me was searching for a validation set that I could rely on. I eventually ended up with a group for tablets from the train set, AKT 8, AKT 12, AKT 6e, and two recent books by Cécile Michel, all restricted to sentence-level pairs. Of these, the tablets that seemed to be closest to the test set were the texts in Michel's \"Women of Aššur and Kaneš - Texts from the Archives of Assyrian Merchants\" and, ironically, the ones furthest from the test set were the texts from train.csv. But there were still significant gaps between my valid and test scores. My best model scored 42.1 valid, 36.3 public LB, 38.1 private LB.\n\nOne technique I used, which I haven't seen mentioned yet, was to include each sentence's starting line number as well as surrounding text (before and after each sentence ) as additional context in the prompt. \n\nLooking at the solutions published so far, I can see I missed a number of opportunities to improve my solution. Congratulations to the winners!" } }