datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
cmbench-0908newvideo
CMBench 0908newvideo
Eight one-minute real first-person videos with clip prompts, continuation (benchmark) prompts, and
keyframe + bounding-box annotations, built for evaluating long-video context memory: whether a
video generation model can bring back an object or a view it saw earlier in the same video.
The metadata format and prompt style follow the earlier CMBench integrated set
(metadata_clip_and_gen_rewrite_20260904.json) and the H3 natural memory benchmark recipe.… See the full description on the dataset page: https://huggingface.co/datasets/Aoraku/cmbench-0908newvideo.CMBench_v2_10subset
CMBench priority-10 videos and prompts (2026-08-19)
This package contains the ten priority context videos (one 60-second source video per priority bundle) and the exact 26 only-gen benchmark prompts used in the 2026-08-18/19 LingBotWorld run. The prompt registry records whether each question is object_recreate or rotate, its source clip id, and the prompt text.
The run configuration was seed=1, Ours probe step/timestep 2, threshold 0.05, generation-chunk pruning enabled… See the full description on the dataset page: https://huggingface.co/datasets/Aoraku/CMBench_v2_10subset.CMBench_merge
CMBench merged dataset — 135 cases
This directory merges the 127-case annotation set from 2026-07-27 with the
eight-case Chinese Prompt annotation set from 2026-08-03. The original 127
cases retain IDs cmb_0001 through cmb_0127; the eight appended cases are
renumbered cmb_0128 through cmb_0135.
Layout
videos/cmb_0001.mp4 ... videos/cmb_0135.mp4: self-contained benchmark
videos.
metadata/benchmark_cases.jsonl: canonical line-delimited metadata.… See the full description on the dataset page: https://huggingface.co/datasets/Aoraku/CMBench_merge.