The pre-distillation SFT corpus: same gold + §4 skeleton as vmar-sft-distill but with
deterministic template reasoning prose. Released as the baseline that motivated teacher
distillation (training on these template-thin targets template-collapses the model). Use for
ablations (template vs distilled reasoning).
train 84,293 · val 10,448 · test 10,522 · ~3,763 audio-hours · n_hops 1–30.… See the full description on the dataset page:
https://huggingface.co/datasets/ngqtrung/vmar-sft-seed-v2.
