Hugging Face's logo

Datasets:
OpenSAE
/
SAEBase-Evaluation

Modalities:
Tabular
Text
Formats:
csv
Languages:
English
Size:
10K - 100K
ArXiv:
Tags:
sparse-autoencoder
mechanistic-interpretability
evaluation
reproducibility
benchmark
Libraries:
Datasets
Dask
Polars
Dataset card Data Studio Files
xet
 Community
2
SAEBase-Evaluation / PROVENANCE.md
Hunter26666666's picture
Hunter26666666
Cite SAR and RSS and clarify dataset documentation
08b3bcb verified
preview code
|
Raw History Blame Contribute Delete
1.45 kB

Source and verification

Historical verification (v0.3.0)

  • SAEBase v0.3.0 tag resolves to d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279.
  • Verified population: 18 configurations and 48 layer sites, each with 64 held-out controlled groups.
  • verified_sites.csv records the model, SAE, layer, hook, and four measured metrics for each site.
  • metric_results.csv contains one row for each (site, metric, condition) combination.
  • Every record links to a per-site JSON from the source at the pinned commit.
  • Numeric records were read from the original source JSON. The metrics were not recomputed.
  • Checkpoints are neither included nor redistributed.
  • The original raw evaluation remains in separate checksum-verified GitHub releases.

The v0.3.0 verification establishes engineering acceptance under controlled inputs. It does not establish quality rankings across models, generalization to representative natural-language data, or certification of every checkpoint.

Research Lab measurements (v0.4.0)

The separately reported 500-prompt Research Lab experiment has its own input population and measurement protocols. See the v0.4.0 Research Lab report for the experiment's source records, limitations, and instructions for verification. The v0.3.0 verification counts above should not be combined with the v0.4.0 research measurements.