Immutable public release
gb2-20261111-7b598bb13de4
DuelLab GameBench 2 evaluates AI-generated game-playing programs through deterministic head-to-head matches. Match outcomes, not human or AI judges, produce the ratings. Public standings, methodology, and release data are available; the evaluator, engine implementation, and unpublished operational artifacts remain private.
- Generated
- Benchmark version
- 2-known8-r3
- Games
- 8
- Models
- 105
- Reasoning settings
- 284
- Rated matches
- 627550
GameBench 1 and GameBench 2 use different game sets and evaluation contexts; their scores and ranks are not directly comparable.
Data and checksums
Leaderboard JSON index (rows in per-leaderboard files under leaderboard/) · Canonical CSV · CSV for spreadsheets · Leaderboard Parquet · Player programs JSON · Charts data · Data dictionary (as of this release)
Release manifest and SHA-256 checksums · Licensed under CC BY 4.0.
Release lineage and package-ID mapping explains when multiple immutable package IDs contain the same canonical result data.
A living copy of the public results is available from the DuelLab GameBench 2 results mirror on Hugging Face. This immutable DuelLab release remains canonical.
Scripts must send their own User-Agent header: Python's default one (also used by pandas for a URL) gets HTTP 403. The data dictionary shows how.
Methodology and definitions
Methodology (as of this release) · Data dictionary (as of this release)
These copies were frozen when this release was published. The current methodology and current data dictionary describe the latest release and can differ.
Citation files
BibTeX · CSL-JSON · RIS · CITATION.cff · Plain text
The canonical citation authority is https://duellab.org/citation.