A benchmark dataset for natural-language recommendation retrieval: given a free-text user request, retrieve the relevant items from a 94 k movie & TV catalog. Prompts are LLM-generated in 12 writing styles grounded in real user interaction histories. Pre-computed 1024-d embeddings (Yuan-embedding-2.0-en) are included for both items and prompts.
