Universal Source Modeling Challenge

This is a live sequential prediction competition: your predictor assigns probabilities to the next symbol using only past symbols, and the score is average log-loss in bits per symbol.

Download the latest release ZIP from GitHub Releases. Unzip it, read the README inside, edit template_predictor.py, and run the dry run.

Quick Start

Smoke test:

python run_live_eval.py --test-path dry_run/release/test.npy --predictor-path template_predictor.py --smoke-test

Dry run:

python run_live_eval.py --test-path dry_run/release/test.npy --predictor-path template_predictor.py

Final competition day uses a different hidden released test file than the dry run, with a different seed, hash, and run ID.