Grounded multi-hop audio-reasoning SFT corpus for MERaLiON-3. Each row is an audio
question with a verifiable §4 reasoning target: a
chain + one per
reasoning step with a time-span , ending in .
Distillation, hallucination-proof by construction: a teacher LLM (Qwen3.6-27B) writes
ONLY the prose ( + per-hop ); the gold caption events, every
span, the … See the full description on the dataset page: https://huggingface.co/datasets/ngqtrung/vmar-sft-distill.
