Datasets:
Accept the Sumeng Video Dataset Terms
Submit this form to request access. The dataset owner manually reviews every request, and access is granted only after approval.
By requesting access, you agree to use this dataset only for lawful research, education, evaluation, or model development; comply with applicable copyright, privacy, publicity-right, and data-protection laws; not redistribute the dataset or use it for biometric identification, face recognition, or surveillance; not attempt to identify depicted persons; apply reasonable security controls; and accept that the dataset is provided as-is without warranties. Access may be revoked for misuse. Source material may be subject to third-party rights, and access does not grant any rights to that material.
Log in or Sign Up to review the conditions and access this dataset content.
Chinese Drama Multimodal 43K
A multimodal video-training corpus produced by the Sumeng data_tool pipeline.
The snapshot contains 43,578 subtitle-cleaned clips from 1,211 episodes across
37 source groups, together with English/Chinese metadata and aligned control
signals used for LTX-Video fine-tuning.
Access
This is a manually gated public dataset. A signed-in user must accept the Dataset Terms, submit the requested account information, and wait for the dataset owner to approve the request. Dataset files are not downloadable before manual approval.
Snapshot
| Field | Value |
|---|---|
| Snapshot date | 2026-08-03 |
| Clips | 43,578 |
| Episode archives | 1,211 |
| Source groups | 37 |
| Unpacked size | 1,416,307,968,370 bytes |
| Original files | 2,761,713 |
Each clip is typically 6-10 seconds at 24 fps. Most video clips are 1280x544 or 1280x704 H.264 MP4 files after subtitle removal and quality filtering.
Contents
Each data/<tv-hash>/<episode>.tar archive preserves one complete episode
directory from the extraction snapshot:
tv-<hash>/<episode>/
├── clean_shots/ # subtitle-cleaned target video clips
├── audio/ # aligned WAV audio
├── camera/ # 6-DoF camera estimates
├── canny/ # Canny control videos
├── depth/ # depth control videos
├── hands/ # hand landmarks
├── images/ # extracted/key frames
├── masks/ # segmentation masks
├── pose/ # rendered pose controls and keypoints
├── pose3d/ # 3D pose where available
├── records/ # clips.jsonl and per-clip records
├── refs/ # scene, face, and person references
└── tracks/ # point/object tracks
Top-level corpus indexes are stored in metadata/. metadata/shards.jsonl
records each archive's path, byte size, SHA-256 digest, and clip count.
Record schema
Each line in an episode's records/clips.jsonl is an enriched clip record with
these top-level groups:
audio, cached, camera, caption_structured, characters,
chinese_drama, color, composition, controls, expression, metadata,
motion, pose, prompt, prompt_raw, props, quality_flags, refs,
scene, source, timeseries, video, clip_id, and version.
Some path values preserve the extraction server's original absolute paths.
After extracting an episode archive, rebase /data0/dataset/v1/ to the local
directory that contains the extracted tv-<hash>/ tree.
Example
Download requires a Hugging Face account that has accepted the gate:
hf auth login
hf download SyFeee/chinese-drama-multimodal-43k \
data/tv-<hash>/<episode>.tar \
--repo-type dataset \
--local-dir ./sumeng-data
tar -xf ./sumeng-data/data/tv-<hash>/<episode>.tar
Intended uses
- Video-generation research and model development
- Caption, camera, motion, pose, depth, segmentation, and identity-control studies
- Evaluation of multimodal video understanding and conditioning systems
Users are responsible for determining whether their intended use is lawful and consistent with the Dataset Terms. The dataset is provided without warranties, and access does not convey rights to any third-party source material.
Attribution
Dataset processing and curation: SyFe / Sumeng Video.
- Downloads last month
- 345