Checklist EUROAI (CEOAI) 2026 Contest Day 2 · Task 2
Interleaved Lines
Separate two randomly interleaved token sequences by predicting which paragraph each token came from.
The task
A synchronisation bug interleaved the tokens of two paragraphs A and B into one sequence C, preserving the order within each paragraph. The contestant predicts a binary mask over C (0 for A, 1 for B); a mask and its inverse are equally correct.
Text is tokenised with the Pythia-14M tokenizer. train_data.csv has 5,000 examples (id, c, mask) and test_data.csv has 500 examples (ids 5000–5499) with only id and c. The starter archive contains the pythia-14m weights and tokenizer loadable with Hugging Face AutoTokenizer and AutoModelForCausalLM.
Abridged by SOTA from the official materials. The official statement has the exact rules, and it wins wherever this summary differs.
At a glance
- You get
train_data.csv(5,000 rows),test_data.csv(500 rows),custom_archive.zip(starter.ipynb, pythia-14m model and tokenizer).- You submit
- CSV with subtaskID (always 1), datapointID (test id) and answer (the comma-separated mask).
- Scoring
- Token-level accuracy per example, taking the better of the mask and its inversion, averaged over the test set. Points: 5 if score < 0.60, 100 if > 0.81, otherwise 5 + round(95 × (accuracy − 0.60) / 0.21).
- Rules
- Submission limit 50, one final submission (platform settings).
- Format
- Contest Day 2, 17 July 2026 (Nitro AI Judge 07:10–12:50 UTC), on-site in Cluj-Napoca; three tasks, 100 points each.