ludwigw/causal-reasoning-benchmarks

--- license: mit task_categories: - text-classification language: - en tags: - causal-reasoning - transitivity - d-separation - axiomatic-training - semantic-loss pretty_name: Causal Reasoning Benchmarks size_categories: - 10K<n<100K --- # 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: ```json {"premise": "A causes B. B causes C.", "hypothesis": "Does A cause C?", "label": "Yes"} ``` ## Citation ```bibtex @article{deshmukh2026semantic, title={On Semantic Loss Fine-Tuning Approach for Preventing Model Collapse in Causal Reasoning}, author={Deshmukh, Pratik and Gupta, Atirek}, year={2026} } ```