Discord

Checklist IOAI 2026 Individual Contest · Task 5

Ghost of the Machine

Find the character position at which each English passage switches from human-written text to a language-model continuation.

  • NLP
  • Boundary detection (text changepoint)

The task

The story is that books in a national archive have been quietly altered. Each sample is one English passage that begins as human writing and, at a hidden point, carries on with text a language model generated from the human part. Read as a whole, it seems like one coherent piece.

Every passage has exactly one switch. Each side has at least 180 words, and passages total about 500 to 800 words. You predict the character offset where the human part ends: everything before it is human, and everything from it onward is machine-written.

Scoring rewards getting close, not only being exact, since each sample's score decays exponentially with the distance in characters. The only pretrained weights allowed are those of the bge-base-en-v1.5 text encoder, which may be used frozen or fine-tuned. Classical feature-based models are unrestricted.

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
JSONL passages ({"id", "text"}): dataset/train has 1,221 samples with answers.jsonl, and dataset/test_public has 380 samples with a development copy of answers. At grading time test_public is replaced by hidden sets of 380 samples without answers: test_leaderboard_a for the public leaderboard and test_leaderboard_b for the final ranking.
You submit
answers.jsonl at the repository root, with one {"id", "boundary_char_index"} object per sample; the index must be an integer in [0, len(text)], and a missing, non-integer or out-of-range entry scores 0 for that sample. Submit a single notebook named exactly solution.ipynb.
Scoring
Per sample exp(−|p − t| / τ) with τ = 100, where p is the predicted and t the true boundary character index; the final score is the mean over the split on a 0–100 scale.
Rules
  • Time limit 10 minutes wall-clock, covering any grade-time training or fine-tuning plus inference
  • One GPU (≈16 GB VRAM), no internet, 5 GB storage
  • solution.ipynb ≤ 20 MB
  • Only pretrained model: bge-base-en-v1.5 (110M-parameter text encoder), frozen or fine-tuned on train; classical and statistical tools are unrestricted
Format
Individual Contest, Day 2 (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

Details

Year
2026, Astana, Kazakhstan
Rounds
Individual Contest · Task 5 · Day 2GAITE Contest · Task 5 · Day 2
Language
English
License
CC BY 4.0, as stated by the source