Anonymous47621123/AgenticVBench_100

AgenticVBench — 100 video-production tasks

AgenticVBench is a benchmark of 100 video-production tasks for evaluating multimodal AI agents. Tasks span four families of the post-production workflow:

family	task_id range	rows	what the agent does
recap	1–36	36	cut a short creative recap of a film/broadcast to a brief written by a domain expert
sequencing	37–64	28	reassemble shuffled chapter shots into the original narrative order
repair	65–82	18	identify and fix one or more defects in a corrupted video
assembly	83–100	18	pick one candidate clip per slot to match a cinematic-language storyboard

Each task family is evaluated with its own scoring rubric; the per-task spec is serialized as JSON in the rubric_json column. Verifier code lives in a separate code repository (linked below).

Schema (6 columns)

column	type	description
task_id	int64	sequential 1..100
task_family	string	recap / sequencing / repair / assembly
prompt	string	the brief shown to the agent
reference_file_urls	list[string]	inputs the agent sees
verifier_reference_urls	list[string]	inputs only the verifier sees (null except for repair)
rubric_json	string	family-specific scoring spec, JSON-serialized

Layout

data/
  train-00000-of-00001.parquet      one row per task (100 total)
recap/materials/<task_id>/film.mp4         source video per recap task
sequencing/materials/<task_id>.zip         candidate clips (zipped)
repair/materials/<task_id>/broken.mp4      the broken video the agent must fix
repair/sources/<variant>_source.mp4        original (verifier-only) video for scoring
assembly/materials/<task_id>.zip           candidate clips (zipped, audio stripped)

Loading

from datasets import load_dataset

ds = load_dataset("Anonymous47621123/AgenticVBench_100", split="train")
print(ds)               # 100 rows
print(set(ds["task_family"]))   # {'recap','sequencing','repair','assembly'}

# Filter to a family
recap = ds.filter(lambda r: r["task_family"] == "recap")

# Parse a task's rubric
import json
spec = json.loads(ds[36]["rubric_json"])  # task_id=37 (sequencing #1)
print(spec["correct_order"])

rubric_json shape per family

• recap: a list of expert-authored rubric items, each with id, weight (signed), dispatch (python / llm / compound), kind, and family-specific parameters. ~28 items per task.
• sequencing: {"correct_order": [...], "n_slots": int, "provenance": {...}}.
• repair: {"cell": "<vN>/<sM>"}. The verifier code's cells.json + ground_truth/<cell>/profile.json provide the rest.
• assembly: {"correct_assembly_in_slot_order": [...], "correct_assembly_durations": [...], "n_slots": int, "film": str, "duration_sec": float}.

Verifier-only data

For the repair family, repair/sources/v<N>_source.mp4 are reference videos used only by the verifier. They are public on this dataset for reproducibility, but agents being evaluated should not see them at rollout time — that would trivialize the task. Honoring this boundary is the responsibility of whoever runs the agent.

Composition

• 100 tasks, English-language, video duration ranging from short clips (~1 s) to multi-minute broadcasts.
• Source videos include short films, broadcast clips, festival shorts, podcast/keynote recordings, and sports broadcasts.
• Annotations were authored by experienced video-production professionals.

Intended use

Evaluating multimodal AI agents on long-horizon video-production tasks that combine planning, tool use, and audio/video understanding. Not intended as training data.

Known limitations

• English-only.
• The recap family's rubrics encode per-expert editorial preferences; high recap scores indicate alignment with expert style, not "correctness" in any absolute sense.
• The repair family's signal-processing rubrics (SSIM, xcorr) are sensitive to encoding choices; agents that re-encode aggressively can underscore even when their fix is perceptually fine.
• Source videos in recap are full-length pieces (multi-minute, multi-GB in some cases). Loading them into long-context multimodal models may exceed individual model context limits.

Licenses

• The curated dataset (prompts, rubrics, correct orderings, ground-truth defect profiles, manifest schemas) is released under CC-BY-4.0.
• The source videos referenced in recap/materials/, sequencing/materials/, assembly/materials/ may carry their own copyrights from the original film/broadcast producers. These clips are included for academic evaluation purposes; users redistributing or building on top of them should verify the underlying license of each source film.
• The broken/source videos in repair/ are derivative works of a curated set of public-domain or permissively-licensed source footage.

Citation

(See the companion paper.)