OMTG-56K Training Dataset

OMTG-56K: A High-Quality Instruction-Tuning Dataset for One-to-Many Temporal Grounding OMTG-56K is a large-scale, high-fidelity instruction-tuning dataset introduced in the paper “Towards One-to-Many Temporal Grounding” (ICML 2026, under review). It empowers MLLMs to evolve from one-to-one to one-to-many temporal grounding via SFT + RL (GRPO). Dataset Summary Item Value Task One-to-Many Temporal Grounding (OMTG) Total samples ~56,000 SFT split ~46… See the full description on the dataset page:

Source: Hugging Face Hub (insomnia7/omtg56k). Metadata imported from the dataset’s Hub tags.