Checklist APOAI 2026 Contest · Task 1
Cosmic Probe: Star–Galaxy–Quasar Classification from SDSS Photometry
Classify SDSS objects as GALAXY, STAR or QSO from five-band photometric magnitudes, with recall on the rare quasar class emphasised by an F2 metric.
The task
Spectroscopic follow-up of sky-survey sources is expensive, so machine learning is used to classify objects from photometry. The contestant acts as a 'digital astronomer' and builds a model that uses the five-band model magnitudes (u, g, r, i, z) to classify each object as GALAXY, STAR or QSO (quasar). Quasars make up only about 18.8% of the data and are easily confused with stars or compact galaxies.
The data come from SDSS Data Release 17 public catalogues (PhotoPrimary joined with SpecObj): 60,000 records split into train.csv (30,000 labelled), valdata.csv (15,000, public leaderboard A) and testdata.csv (15,000, private leaderboard B). Columns are objid, ra, dec, modelMag_u, modelMag_g, modelMag_r, modelMag_i, modelMag_z and, in training data, type. Class proportions are preserved in all splits.
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
train.csv(30,000 rows with features and label);valdata.csvandtestdata.csv(15,000 rows each, features only). Two reference papers (PDF) are included with the training data.- You submit
submission_val.csvandsubmission_test.csvwith columns objid and type, in the same row order and length as the corresponding test file; submitted as a notebook on Bohrium.- Scoring
- F2 score (β = 2), computed per class and aggregated as a weighted average by class support; final ranking uses leaderboard B only. Baseline on B: 0.6660; scientific committee reference: 0.8181.
- Rules
- Pretrained model weights are not allowed, including weights pretrained on ImageNet or any other external dataset.
- External large language model APIs (e.g. GPT, Claude) may not be used for prediction, feature generation, data labelling or model ensembling.
- Single training run of the notebook ≤ 20 minutes on CPU.
- Evaluated in the 'ioai3.6' image.
- Contestants may read the referenced papers but may not use pretrained models, released weights or external code repositories associated with them.
- 15 submissions per day in the practice re-run.
- Format
- APOAI 2026 contest (first edition): a six-hour examination on Saturday 27 June 2026, taken online on the Bohrium platform with local invigilated sites; the four tasks were republished on Bohrium for practice from 15 June to 31 July 2026 (platform time UTC+8).