ChaosBench-Logic / docs /DATASET.md
11NOel11's picture
Upload docs/DATASET.md with huggingface_hub
6c8d241 verified

ChaosBench-Logic Dataset (v2.0.0)

This document is the canonical in-repo reference for the released v2 dataset.

Release Identity

Field Value
Dataset release v2.0.0
v2 questions 40,886
v1 archived questions 621
Total questions 41,507
Canonical files 10 (data/v22_*.jsonl)
Selector file data/canonical_v2_files.json
Manifest data/v2_manifest.json
Global SHA256 cfcfcc739988ad99c38d47dd171ff39f67df3ddca7d8d452e8c77b30f14e7279

Canonical v2 Files

File Count
data/v22_atomic.jsonl 25,000
data/v22_multi_hop.jsonl 6,000
data/v22_consistency_paraphrase.jsonl 4,139
data/v22_perturbation_robustness.jsonl 1,994
data/v22_adversarial.jsonl 1,285
data/v22_fol_inference.jsonl 1,758
data/v22_indicator_diagnostics.jsonl 530
data/v22_regime_transition.jsonl 68
data/v22_cross_indicator.jsonl 67
data/v22_extended_systems.jsonl 45

Schema

Each JSONL record uses this schema:

Field Type Required Notes
id string yes unique question identifier
question string yes natural language prompt
ground_truth string yes TRUE or FALSE for v2
type string yes task family label
system_id string or null yes null for system-agnostic ontology/FOL items
template string yes template version (for v2, typically V2)

Coverage

  • Systems: 165 total (30 manually curated core + 135 imported from dysts).
  • Predicates: 27 logical predicates used in ontology and reasoning chains.
  • Task families: 10 families spanning atomic, multi-hop, adversarial, indicator-based, and robustness reasoning.

Archive Policy

  • v2 files (v22_*.jsonl) are the default evaluation target.
  • v1 is preserved at data/archive/v1/ for historical baseline comparisons only.

Validation and Reproducibility

  • Manifest and per-file hashes: data/v2_manifest.json.
  • Canonical file selector for tooling: data/canonical_v2_files.json.
  • Freeze artifact generation: uv run python scripts/freeze_v2_dataset.py.
  • Strict validation command: uv run python scripts/validate_v2.py --strict --max-duplicate-questions 200.

External Publication

  • Repository: https://github.com/11NOel11/ChaosBench-Logic
  • Hugging Face dataset: https://huggingface.co/datasets/11NOel11/ChaosBench-Logic

For expanded narrative metadata used in external publishing workflows, see ../DATASET_CARD.md.