mohantesting/video-quality-scored

Image-to-Video Quality-Scored Clips

A collection of prompted image-to-video samples with quality-evaluation metadata. Each sample pairs a first frame (the I2V conditioning image) with one or both of:

• a generated video produced by a video model from the first frame + prompt
• an original clip (the reference/source video the prompt was authored around)

A subset of the samples also carry per-clip quality scores: an overall quality_score, six per-aspect breakdowns (first_frame_fidelity, prompt_adherence, temporal_consistency, motion_quality, visual_quality, camera_composition), and an array of multi-evaluator critiques (evaluations) — free-form reasoning paragraphs paired with structured score breakdowns.

Structure

.
├── metadata/
│   ├── slice_400.jsonl       (smallest preset — 400 rows)
│   ├── slice_500.jsonl
│   ├── slice_600.jsonl
│   ├── slice_700.jsonl
│   ├── slice_1000.jsonl
│   ├── slice_2000.jsonl
│   └── slice_4000.jsonl      (full superset — DEFAULT)
└── media/
    ├── sample_NNNNNN_first_frame.png       (conditioning image; always present)
    ├── sample_NNNNNN_generated.mp4         (model output; present when available)
    └── sample_NNNNNN_original_clip.mp4     (reference clip; present when available)

All seven slice metadata files share the same media/ directory. Slice budgets are nested: slice_400 ⊂ slice_500 ⊂ ... ⊂ slice_4000. A given sample_id refers to the same underlying clip across every slice it appears in.

Loading

from datasets import load_dataset

# Default (slice_4000 — full superset, 4000 rows)
ds = load_dataset("mohantesting/video-quality-scored")

# Pick a smaller preset
ds_small = load_dataset("mohantesting/video-quality-scored", "slice_500")
ds_mid   = load_dataset("mohantesting/video-quality-scored", "slice_1000")

To resolve a media path:

from pathlib import Path
row = ds["train"][0]
first_frame = Path(ds.cache_files[0]["filename"]).parent / row["first_frame"]

Schema

Each row of slice_*.jsonl:

Field	Type	Notes
sample_id	str	Zero-padded 6-digit string. Stable across slices.
prompt	str	Natural-language description of the desired video.
first_frame	str	Relative path to the conditioning image.
generated_video	str | null	Relative path to the model-generated clip. Null when no generation is available.
original_clip	str | null	Relative path to the reference/source clip.
quality_score	float | null	Aggregate quality score (0–100), null when no evaluation was captured.
first_frame_fidelity	float | null	Per-aspect score.
prompt_adherence	float | null	Per-aspect score.
temporal_consistency	float | null	Per-aspect score.
motion_quality	float | null	Per-aspect score.
visual_quality	float | null	Per-aspect score.
camera_composition	float | null	Per-aspect score.
passed_quality_threshold	bool	quality_score >= 80.0. False when score is null.
evaluations	list	Per-evaluator critiques (see below). May be empty.

Each entry of evaluations:

Field	Type	Notes
evaluator_id	str	Anonymous identifier (eval_0, eval_1, ...). Stable within the row.
passed	bool	Whether the evaluator marked this clip as passing.
confidence	int	Evaluator self-reported confidence (0–100).
evaluated_at	str	ISO 8601 timestamp.
reasoning	str	Free-form rationale (typically 1–3 sentences).
aspect_scores	dict	Six per-aspect scores (same keys as the row-level breakdown).
overall_score	float	This evaluator's overall score.

Slice statistics

Slice	Rows	Has generated_video	Has quality_score
slice_400	400	342	342
slice_500	500	424	424
slice_600	600	517	517
slice_700	700	603	603
slice_1000	1000	850	850
slice_2000	2000	1679	1679
slice_4000	4000	3372	3372

Notes

• Resolution & length. Most clips are ~832×480 at ~5 seconds. Some samples are slightly off-spec. Re-encode or resize as needed.
• Quality scores range 0–100; the convention used here treats 80 as a practical "passing" threshold, reflected by passed_quality_threshold.
• Multi-evaluator coverage. Rows with evaluations typically carry 3–4 independent critiques. The aggregate quality_score and aspect scores are computed from these (mean over evaluators).
• Reference vs generation. Many samples carry both an original_clip (reference truth) and a generated_video (model output) — useful for comparative/reward-model training. A subset carries only the reference clip.

License

CC-BY-4.0 — please cite this dataset if you find it useful.