Checklist IOAI 2026 At-Home Round · Task 1
Operation Night Watch
Extend a deployed 16-class Audio Spectrogram Transformer to recognise 29 sound classes without forgetting the original ones.
The task
An Audio Spectrogram Transformer (AST) running on acoustic monitoring stations recognises 16 everyday and wildlife sound classes. Two new customers need 13 more: nine sounds tied to illegal logging and poaching, such as chainsaws, gunshots, helicopters and crackling fire, plus the calls of four singing insect species.
You start from the provided 16-class checkpoint, a small retained subset of the old training data and new data for the 13 added classes. The new data is imbalanced, with 24 to 60 training clips per class. The upgraded model must classify all 29 classes in a single forward pass.
Retraining from scratch is not allowed. You must reuse the checkpoint's encoder weights, though you may add parameters and fine-tune any subset of weights, and no other pretrained audio model may be used. Old and new classes count equally in the score, so forgetting the original sounds costs as much as failing to learn the new ones.
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
- An AST checkpoint (model/, 16 classes),
train.csv(retained old data, 16 classes, train/val splits),fine_tune.csv(new data, 13 classes, train/val splits) and 5-second mono 16 kHz .wav clips, downloaded asdataset.zipfrom Google Drive. The notebook's saved outputs show 491 old-data clips and 792 new-data clips. - You submit
- A 29-class model evaluated on a hidden held-out test set covering all 29 classes; the notebook does not state a submission file format.
- Scoring
- Score = ½ · Acc_old + ½ · Acc_new, where Acc_old and Acc_new are accuracies over test clips whose true label is one of the 16 old or 13 new classes, in a single 29-way classification on a held-out test set.
- Rules
- Must start from the provided checkpoint and reuse its encoder weights (adding parameters and fine-tuning any subset of weights are allowed)
- The final model must classify all 29 classes in a single forward pass
- No training from scratch and no other pretrained audio models
- Lightweight: a single GPU and about 10 minutes of training time
- Format
- At-Home Round (educational; the 2026 Contest Rules say the three problems were given to teams about one month before IOAI 2026 and do not affect final results)