Checklist IOAI 2026 Individual Contest · Task 1
Find the Order
Recover the original chronological order of shuffled speaker-turn audio clips from two-person spoken English dialogues.
The task
Two-person spoken English conversations have been cut into speaker turns, one .wav file per turn, and the files have been shuffled. A name such as chunk_3.wav only reflects the shuffled position, not where that turn really came in the conversation.
Each dialogue has 7 to 20 mono, 44.1 kHz chunks. A prefix.json file names the chunks that are the first and second turns, which fixes where the dialogue starts and rules out reading it backwards. You get a labelled train folder of 1,288 dialogues and a test_public folder of 100 dialogues for scoring yourself locally.
For every dialogue you output a permutation giving each chunk its true position in time, saved as answers.json. At grading time the notebook is run again on hidden sets of the same size and format. Only three pretrained model families are allowed: wav2vec 2.0, Whisper (any size, with its encoder usable as a feature extractor) and Qwen2.5-0.5B.
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
- Per-dialogue folders of
chunk_{k}.wavfiles (7–20 per dialogue, mono, 44.1 kHz) plusprefix.jsongiving the first two chunks;dataset/trainhas 1,288 dialogues withanswers.jsonanddataset/test_publichas 100 dialogues with answers. At grading timetest_publicis replaced by hidden sets of the same size and format:test_leaderboard_afor the public leaderboard andtest_leaderboard_bfor the final ranking. - You submit
answers.jsonmapping each dialogue ID to a permutation P of {0, …, n−1}, where P[i] is the predicted chronological position ofchunk_i.wav(0 = first). An invalid permutation or a missing dialogue scores 0 for that dialogue, and a malformed file is rejected. Submit exactly one file,solution.ipynb, which writesanswers.json.- Scoring
- Pairwise ordering accuracy: per dialogue, score = 1 − I/M, where I is the number of chunk pairs ordered differently from the ground truth and M = n(n−1)/2; the final score is the average over all dialogues in the split.
- Rules
- Time limit 10 minutes, covering any training or fine-tuning done at grade time plus inference
- One GPU (≈16 GB VRAM), no internet, 5 GB storage
solution.ipynb≤ 1 MB- Only these pretrained models: wav2vec 2.0, OpenAI Whisper (any size; the encoder may be used as a feature extractor) and Qwen2.5-0.5B (zero-shot or fine-tuned on train)
- Format
- Individual Contest, Day 1 (the 2026 Contest Rules give six hours for the day's three tasks); the same task was set in the GAITE Contest with a hint