This is a practice mirror of Task 1 from the IOAI 2026 Individual Contest. The original task and data were created by the IOAI 2026 Scientific Committee and are published under CC BY 4.0. This Kaggle competition adapts that task, providing GPU-enabled Notebooks for participants.
You are given spoken English dialogues between two participants, Speaker A and Speaker B. Each dialogue is segmented into speaker turns, with each turn containing speech from only one speaker. Every turn is stored as a separate .wav audio file. Unfortunately, these turns have been randomly shuffled, so the conversation no longer makes sense. Your task is to reconstruct the original chronological order of every conversation.
Each dialogue contains n audio files named chunk_0.wav, chunk_1.wav, …, chunk_{n-1}.wav. These chunks are individual speaker turns, and their filenames describe only the shuffled order. Each dialogue has 7–20 mono audio chunks sampled at 44.1 kHz. For every dialogue, prefix.json contains the filename indexes of the first two true chunks, which identifies the beginning of the dialogue and removes ambiguity regarding forward or backward reading.
The competition provides labeled training dialogues and hidden-label test dialogues:
- Training Data: 1,288 labelled training dialogues including train/prefix.json and train_metadata.csv for training, fine-tuning, and validation.
- Test Data: Hidden-label test dialogues including test/prefix.json, test_pairs.csv, and a copy of sample_submission.csv for generating predictions.
Kaggle evaluates ordering decisions in pairwise form. You must submit a CSV file named submission.csv with exactly two columns: pair_id and before.
- pair_id: Must match the IDs in sample_submission.csv exactly.
- before: An integer value where 1 means chunk_i.wav occurs before chunk_j.wav, and 0 means chunk_j.wav occurs before chunk_i.wav.
For solutions comparable to the original IOAI task, use only the following pre-trained model families: - Speech representations: wav2vec 2.0 (Whisper encoder may also be used as a feature extractor). - Automatic speech recognition: OpenAI Whisper (any size). - Language model: Qwen2.5-0.5B, used either zero-shot or fine-tuned on the training split.
The original task uses pairwise ordering accuracy. A pair is correct when the predicted ordering agrees with the ground truth. While the original IOAI score averages per-dialogue scores, this Kaggle mirror uses standard Accuracy over all rows of submission.csv, meaning it micro-averages all pairwise decisions.
Because the official archival dataset is publicly available, this mirror is intended for practice, teaching, and pipeline validation rather than secure ranking.