Checklist NEOAI 2026 Day 1 · Task 2
Day 1 NLP (Message Author Re-identification)
Using two damaged checkpoints of the same transformer encoder, retrieve for each test message the five messages most likely written by the same user.
The task
Two similar checkpoints of the same transformer model are provided; the model was trained on a similar subset of data that does not include the train and test sets. In the checkpoint "coca_cola_zero" some weights have been randomly set to zero, but the transformer blocks are in the correct order. In "coca_cola_shuffle" a smaller share of weights has been zeroed, making it the higher-quality checkpoint, but its transformer blocks have been shuffled.
The users in the train and test sets do not overlap; this is a re-identification task. For every test message the contestant must find the five most relevant test messages, where a message counts as relevant if it was written by the same user.
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_.csv(text, target = user) andtest_.csv(text); the two damaged checkpoints (coca_cola_zero,coca_cola_shuffle), a model config and a tokenizer. The baseline builds a USER2-base-style encoder (deepvk/USER2-base configuration) with a cosine classification head.- You submit
- CSV with columns index, target, where target is a space-separated list of 5 row indices of test messages.
- Scoring
- mAP@5 (a recommended message is relevant if it was written by the same user).
- Rules
- General NEOAI 2026 rules apply: by default no pretrained models or additional data unless the statement allows them.
- Format
- NEOAI 2026 main round, Day 1 (3 May 2026, 07:00–13:00 UTC). Individual.