Discord

Checklist NEOAI 2025 Day 2 · Task 5

Broken BERT

Repair the corrupted token-embedding matrix of a BERT sentiment classifier without training the model.

  • NLP
  • Model repair (sentiment classification)
Read the official statement

The task

A BERT sentiment-analysis model (Hugging Face: Ilseyar-kfu/broken_bert) had its token embeddings damaged; the attention blocks and classification heads are intact. The contestant must fix the broken vectors of the embedding matrix so that the model classifies text sentiment (neutral, positive, negative) better.

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 (text, labels) and test.csv (id, text); the broken model and tokenizer.
You submit
CSV with columns labels, id, produced by the supplied create_submission function.
Scoring
Accuracy.
Rules
  • No other transformer-based pretrained models or LLMs.
  • No additional data.
  • No fine-tuning or pre-training of the model.
  • The task must be solved on Kaggle (cloud.ru may not be used); the notebook must be Quick-Saved and its version cited in the submission comment.
  • General NEOAI 2025 rules apply: by default no pretrained models or additional data unless the statement allows them; no external human help; on-site participants get one NVIDIA V100, an LLM chat (GPT-4o or similar) and a whitelist of websites.
Format
NEOAI 2025 main round, Task 5, Day 2 (8 May 2025). Individual; 26 submissions per day.

Notebooks

Model

Competition

Solutions

Source files

Details

Year
2025, SberUniversity, Russia and online
Round
Day 2 · Task 5
Language
English
License
Per-competition Kaggle data licence: Subject to Competition Rules (Task 4); MIT (Task 5); CC BY-NC-SA 4.0 (Task 6), as stated by the source