siddharthmb/2026.RA.Fairness-Counterfactual-Pairs

2026.RA.Fairness-Counterfactual-Pairs Action-level contrastive pairs from five-party private-information negotiations: at every turn an LLM took, what a computable ideal agent would have done at that same decision point. Each row is one (episode, turn, counterfactual_type). rejected_action is what the LLM actually did; chosen_action is the counterfactual agent's action. Both are structured actions ({atype, offer_id, deal}), never prose. 72,192 rows, 58,168 of them (81%)… See the full description on the dataset page: https://huggingface.co/datasets/siddharthmb/2026.RA.Fairness-Counterfactual-Pairs.