Supersedes the original small version. This is the data the published
BERT / RoBERTa / T5 models were actually fine-tuned on: a label-stratified
75/25 split with seed 42, giving 3,000 train / 1,000 test. The test split
is the one all adversarial attack results are measured on.
Generated with mdg/synthetic/generate_causal.py (task 2, seed 42, n=4000).
