spurious_sentience

Software and data underlying the publication: "Position: Stop Making Unscientific AGI Performance Claims"

4
contributors

Description

Code and research results for ICML 2024 position paper. Originally released here: https://github.com/pat-alt/spurious_sentience.

The research results include:

Regression tables (.tex; .html)An "evaluations.csv" file that contains estimated evaluation metrics for linear probes and the baseline grouped by indicator, layer (network layer), train/test split, variable (measure), model (lin. probe/baseline).A figures/ folder containing all PNG figures that went into a) the body or b) the appendix.An interim/ folder containing results for probe predictions for each training epoch.An attacks/ folder containing the CSV files of neural network activations for attack prompts (see paper for details). Additionally, this folder contains a sentences/ subfolder with the actual textual attack prompts (.txt files).

Logo of spurious_sentience
Keywords
Programming language
  • Other 100%
License
  • MIT
</>Source code
Packages
data.4tu.nl
data.4tu.nl

Contributors

AD
Andrew Demetriou
AB
Antony Bartlett
CL
C.C.S. (Cynthia) Liem

Member of community

4TU