Checklist GAIA AI Olympiad 2026 Georgian AI League I (practice contest) · Task 2
Planet X Model Selection
Georgian title: პლანეტა X-ის მოდელის შერჩევა
From the training-set predictions of 1,000 regression models, identify the 5 models that were trained on uncorrupted labels.
The task
Scientists on Planet X tried to predict a phenomenon Y from X features, but gravitational anomalies sometimes damaged their instruments. Over the years they trained 1,000 regression models, of which exactly 5 were trained while the instruments worked correctly; the rest were trained on corrupted labels.
The contestant receives a table whose first 15 columns are the features the models were trained on and whose remaining 1,000 columns are the 1,000 models' predictions on the training data, and must identify the indices of the 5 correctly trained models.
Abridged and translated by SOTA from the official Georgian materials. The official statement has the exact rules, and it wins wherever this summary differs.
In English
This task was published in Georgian. SOTA translated it into English on 17 September 2026.
Read the task statement in English
Planet X Model Selection
English translation by SOTA – AI Community of the Georgian original. Organisers who would like this translation removed can email [email protected].
Source: Georgian AI League I, a practice contest of the Georgian Artificial Intelligence Association (GAIA) on the Nitro AI judge, 18 January 2026, task 2: original statement.
🪐 Task: Planet X Model Selection
Overview
On planet X, scientists were trying to predict the mysterious phenomenon Y on the basis of X features. However, planet X has unusual gravitational anomalies that sometimes damage their measuring instruments during data collection!
Over the years they collected data and trained 1000 regression models. Unfortunately, only a small number of these models (exactly 5) were trained while the instruments were working correctly. The rest were trained on corrupted labels!
Task
You are given:
- Tabular data (
train_data.csv)
Your goal: identify the indices of the 5 correctly trained models.
Input Files
train_data.csv
Tabular data containing:
- the first 15 columns — the features on which the models were trained
- the remaining 1000 columns — the predictions of the 1000 models on the training data
You can load the data as follows:
import pandas as pd
data = pd.read_csv("train_data.csv")
features = data.iloc[:, :15]
predictions = data.iloc[:, 15:]
Output Format
You must create a CSV file containing exactly three columns and 1 row:
subtaskID,datapointID,answer
1,0,"1,4,3,16,246"
where:
subtaskIDis exactly 1 (because of the platform's format)datapointIDis exactly 0answer— a string containing the comma-separated indices of the 5 models (0-999)
Under no circumstances change the column names.
Evaluation
Your work will be scored by accuracy.
Scoring Rules
The final score is given on a 100-point scale. Accordingly, if you correctly identify:
- 1 model, you get 20 points
- 2 models, you get 40 points
- 3 models, you get 60 points
- 4 models, you get 80 points
- 5 models, you get 100 points
Good luck, Earthling scientist! 🚀
Translated by SOTA. The Georgian original is the official version and wins wherever the two differ. Georgian AI League I was a practice contest of the Georgian Artificial Intelligence Association (GAIA) on the Nitro AI judge; the statement exists only in Georgian, and the data can be downloaded there after a free login. If you organise this olympiad and would like the translation removed, email [email protected] and we will take it down.
At a glance
- You get
train_data.csv: 15 feature columns followed by 1,000 prediction columns.- You submit
- A CSV with columns subtaskID, datapointID, answer and one row: subtaskID 1, datapointID 0, answer a quoted comma-separated list of 5 model indices (0–999), e.g. "1,4,3,16,246".
- Scoring
- 20 points per correctly identified model (5 correct = 100 points).
- Rules
- Submission limit 25, one final submission (platform settings).
- Format
- Georgian AI League I, a GAIA practice contest on Nitro AI Judge, 18 Jan 2026, 07:00–11:00 UTC (4 hours), individual, online.