ludwigw/causal-reasoning-benchmarks

Causal Reasoning Benchmarks

Datasets used in "On Semantic Loss Fine-Tuning Approach for Preventing Model Collapse in Causal Reasoning" (Deshmukh & Gupta, 2026).

Dataset Structure

• train/transitivity_train.jsonl — 50,000 transitivity training examples
• train/dsep_train.jsonl — 50,000 d-separation training examples
• eval/length_eval.jsonl — 10,000 length generalization examples
• eval/branching_eval.jsonl — 10,000 branching structure examples
• eval/reversed_eval.jsonl — 10,000 reversed edge examples
• eval/shuffled_eval.jsonl — 10,000 shuffled premise examples
• eval/long_names_eval.jsonl — 10,000 long node name examples

Data Format

Each JSONL line:

{"premise": "A causes B. B causes C.", "hypothesis": "Does A cause C?", "label": "Yes"}

Citation

@article{deshmukh2026semantic,
  title={On Semantic Loss Fine-Tuning Approach for Preventing Model Collapse in Causal Reasoning},
  author={Deshmukh, Pratik and Gupta, Atirek},
  journal={arXiv preprint arXiv:2605.05438},
  year={2026}
}