Checklist AICC 2025 Round 1 · Task 2
Autocorrect
Build a CPU-only autocorrect model, trained from scratch, that fixes a user's systematic typing errors.
The task
A student who types quickly on a small phone screen produces many typos and needs an autocorrect model. His postscript hints that his errors follow a somewhat predictable pattern that statistics on the misspelled sentences can reveal, and that some documents are very long.
The training data pair correct reference text with the misspelled version; the test data give only the misspelled text.
Contestants output the corrected text for each test row, in the original order.
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, misspell) andtest.csv(misspell; 2643 samples).- You submit
- A CSV file with no header containing the corrected texts in the same order as
test.csv. - Scoring
- Character Error Rate, CER = (S + D + I) / N (substitutions, deletions and insertions over the number of reference characters); lower is better. The baseline scores 0.0944.
- Rules
- No pretrained models (no APIs, no fine-tuning of open-source models, no model not trained from scratch in the notebook).
- No training on the test samples in any way (including unsupervised learning or statistics over them).
- CPU-only; all 2643 test samples must be evaluated within 250 seconds.
- Individual participation (maximum team size 1); at most 20 submissions per day.
- AICC contest rules (stated on each Kaggle rules page, not enforceable): no use of LLMs for writing code or getting task ideas; no internet use other than official library documentation and the contest platform; no communication with anyone during the contest; clarifications only via the #clarification-requests channel on the AICC Discord server.
- Format
- AICC Round 1 (November 2025), online on Kaggle; Kaggle competition window 15 Nov 2025 (about 18:50 UTC) – 16 Nov 2025 19:00 UTC.