Dataset Viewer
The dataset viewer is not available for this dataset.
Unexpected token '<', "<html> <h"... is not valid JSON

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

CGA-Bench Hugging Face Collection

This dataset repo is a collection index for the nine reviewer-facing CGA-Bench dataset descriptors used in the NeurIPS 2026 E&D submission.

Included configs

  • overview: collection-level summary row spanning the full benchmark release
  • main_corpus: 19,062-episode primary evaluation corpus
  • source_grounded: source-grounded SGSC subset
  • graph_anchored: graph-anchored SGSC subset
  • profile_expanded: profile-expanded SGSC subset
  • auto_expanded: 76,464-episode typed-CwT baseline expansion
  • ranking_subset: legacy 8-model ranking subset
  • bayes_floor: theorem appendix Bayes-floor subset
  • frontier_audit: held-out/frontier audit slice

Repo layout

  • manifests/*.jsonl: HF viewer-friendly manifest row for each config
  • croissant.json: collection-level Croissant metadata
  • croissants/*.json: subset-specific Croissant metadata files
  • cga_bench_anon_v7.3.tar.gz: anonymized release bundle for full reproducibility
  • cga_bench_anon_v7.3.tar.gz.sha256: tarball checksum

Review note

The HF viewer reads the manifests/*.jsonl files. The full benchmark artifacts remain available through the bundled release tarball and the Croissant metadata files.

Downloads last month
25