Seeded train-test splits and sampling scripts for generative NER prompt optimization

Five-seed (2026--2030) deterministic splits over eight NER benchmarks (BC5CDR, CoNLL-2003, five CrossNER subsets, and Few-NERD). For each (seed, dataset) pair, 200 instances form the optimization set visible during prompt search and 500 instances form the held-out test set. Method-specific sub-splits for OPRO, GEPA, PromptAgent, PromptBreeder, and DHOP are derived from the 200 optimization instances. The repository also includes the sampling script that reproduces these splits from the original public HuggingFace and GitHub datasets.