10,000 field-realistic instruction records for practical post-harvest loss reduction.
Raw operator prompts, source-grounded decision envelopes, and Adaption-improved completions.
PostHarvest teaches an assistant to help when the evidence is imperfect and the operating budget is small. Its users do
not speak in benchmark templates. They describe warm corners, mixed farms, rough timestamps, delayed bruising, uneven
drying, suspect… See the full description on the dataset page:
https://huggingface.co/datasets/prathmeshadsod/postharvest.
