metadata
license: other
pretty_name: Open Interpretability exchange inputs
Open Interpretability: inputs of papers and analyses
The files that papers, analyses and scripts of
Open Interpretability read from its
git-ignored exchange/ folder. Examples are SPD checkpoints and training logs, reviewer
prompts and probe sets, and corpus samples. data/exchange-manifest.json in the
repository pins this dataset's revision and each file's SHA-256, and
scripts/exchange.py fetch downloads and checks them.
Text in these files quotes the corpora that the scans read, under their own terms. Examples are pile-10k (bigscience-bloom-rail-1.0, each Pile component under its own) and FineWeb-Edu (ODC-BY). Model outputs and code-generated files are under the repository's MIT licence. Pile corpus files themselves are not included.