Dataset release policy
CogGym dataset access
As more studies are added, CogGym maintains a three-part release policy: 10% of studies are kept fully public, 40% are available under a research-use agreement as the semi-public set, and 50% remain private for held-out evaluation. This split supports open inspection, tooling, and reproducibility while reducing training contamination and preserving a substantial unseen benchmark.
Both releases should be used with the original-study notices preserved, and work using either release should cite CogGym and the relevant source studies.
Public dataset
Fully open: 10 studies, 23 experiments
The fully open CogGym release contains 23 experiments from 10 studies, spanning five CogGym topic groups and text, image, and video modalities. Every included study has at least one experiment independently replicated in the CogGym paper; the release also includes the study's other benchmark-eligible experiments. It contains the EML experiment definitions and stimuli, aggregate and individual human response data, and standalone evaluation code for constructing prompts, running models, scoring responses, and reproducing example analyses.
No access request or signed agreement is required. Because this subset is fully public, it is intended for format inspection, tooling development, and reproducible examples rather than as a held-out benchmark.
View the open dataset on GitHubSemi-public dataset
Licensed access: 50 total studies, 120 experiments
The larger research set includes all 10 public studies plus 40 additional studies. It is available for noncommercial model evaluation and scientific analysis. Access requires reviewing and electronically signing the research-use agreement. The dataset may not be used for model training, commercial use, or redistribution.
Commercial users should contact lanceying@mit.edu.
Review agreement and request access