Why this matters Prompts are becoming a primary interface for generative models, but raw prompts scattered across platforms are hard to reuse systematically. This dataset turns web-sourced Seedance 2.0 video prompts into a machine-readable collection so researchers and engineers can treat prompts as first-class data for analysis, tooling, or downstream model training.
What Sets It Apart
- Unified JSONL schema: each entry standardizes title, bilingual prompt text, tags, media file paths, and granular specs (width, height, ratio, duration, safety_rating), making bulk import into dataframes or databases trivial.
- Media-linked records: entries map directly to cover JPGs and generated MP4s so you can inspect model outputs alongside prompts without manual asset matching.
- License and provenance focus: curated for reuse under CC BY 4.0 and includes source links and basic provenance fields, lowering legal friction for commercial or research use.
- Continuous sync and multilingual coverage: the hub indexes prompts in English and Chinese and is maintained with automated updates, so scale and composition evolve over time.
Who it's for and tradeoffs
Great fit if you are building or evaluating text-to-video pipelines, studying prompt-engineering patterns, or need a labeled corpus of prompts paired with model outputs for fine-tuning or metric analysis. It is less appropriate if you require rigorously curated human annotations, guaranteed copyright clearance for all media, or a stable, immutable snapshot—entries are aggregated from public communities and the dataset is continuously updated, so provenance detail and quality vary across items.
Practical notes
Expect a large, growing download (project pages report differing totals as the collection syncs). Typical workflows: load the metadata.jsonl into pandas or a database, inspect media via the media file paths, and filter by spec or tags for targeted experiments. Due diligence on original creators and media rights is recommended when using generated videos beyond internal research.
