Checklist Cyprus AI Camp 2026 IOAI Selection Camp · Task 7
Broken BERT
Repair the corrupted token-embedding matrix of a BERT sentiment classifier so that it classifies texts correctly again.
The task
A BERT sentiment model (Hugging Face 'Ilseyar-kfu/broken_bert') had its token embeddings corrupted; the attention blocks and classification head are intact. Contestants must fix the broken vectors of the embedding matrix to restore quality on three-class sentiment analysis (negative, neutral, positive).
The baseline notebook loads the model, leaves a slot for replacing model.bert.embeddings.word_embeddings, evaluates macro F1 on validation data and writes submission.csv.
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
val_dataset.csv(labelled texts; 2,500 rows in the baseline run) andtest.csv(texts with ids from 5000).- You submit
submission.csvwith labels (negative/neutral/positive) and id.- Scoring
- Kaggle metric: accuracy (the baseline also reports macro F1 on validation). Points formula not stated on the page.
- Rules
- Pre-trained models and additional data are not allowed unless the task says otherwise.
- No outside help; LLM chats, ChatGPT, Copilot and similar tools are prohibited (except Colab inline autocompletion).
- Only the Kaggle website and Google Colab may be used (no local IDEs); online library documentation is allowed.
- One Kaggle account per participant; leaderboard name '[City] Name Surname'.
- No other transformer-based pre-trained models or LLMs.
- No additional data.
- Format
- Cyprus AI Camp 2026 (IOAI 2026 selection, in person, 31 Jan–8 Feb 2026): NLP, released 6 Feb 2026, 09:00–12:00 EET, on Kaggle (two tasks per 3-hour session).