Supervised fine-tuning data for a small model that, given a high-school physics
flashcard (question + answer), writes one new question testing the same underlying
principle in a clearly different real-world scenario — a far transfer, not a
paraphrase. The model is never told the principle name; it must infer and preserve it.
Example. Original: "What causes atmospheric pressure?" →
Variant: "What causes the water… See the full description on the dataset page:
https://huggingface.co/datasets/meric533/physics-recontext-sft.