OffTrail: A Benchmark for Natural-Language Context in Recommendation beyond Interaction Histories

OffTrail is a benchmark dataset for evaluating whether natural-language situation context can improve recommendation when interaction history alone is weak or ambiguous. Built on MovieLens, the dataset contains fixed recommendation episodes with user histories, target movies, item metadata, generated life-situation contexts, and evaluation artifacts for paired context/no-context comparison. OffTrail includes two complementary scenario families: Observed, where the target is a real user choice that is difficult for a history-only recommender, and Neighbor, where similar recent histories support plausible but divergent next-item continuations. The release is intended as a reproducible benchmark for studying context-aware, sequential, LLM-based, and hybrid recommendation models, while treating generated contexts as plausible external circumstances rather than ground-truth user intentions.