CineSubBench

CineSubBench evaluates long-form narrative and cultural understanding from multilingual movie subtitles. A subtitle track presents a film as thousands of short, temporally ordered utterances. Models must reconstruct characters, relationships, events, causal progression, and themes without explicit speaker, scene, or event structure. The benchmark contains 1,012 films with complete coverage in six subtitle languages and ten national motion-picture rating systems, yielding 6,072 tracks and 8.13 million timestamped entries.

Figure 1. Overview of CineSubBench. The same films support matched multi-task, multilingual, and multicultural evaluation.

Seven tasks span the two groups used in the paper:

Task group	Tasks	What they probe
Narrative understanding and generation	Plot, spoiler-aware synopsis, key message, genre	Premise reconstruction, event selection and chronology, thematic abstraction, and multi-label classification
Cultural prediction and assessment	Age suitability, country-specific motion-picture ratings, language safety	Audience guidance, prediction within ten distinct national label spaces, and subtitle-indexed evidence grounding

The same films support cross-lingual comparison in six subtitle languages. Language-safety evidence is evaluated on the English track; its annotations include category counts and exact subtitle indices.

In the submitted evaluation of nine LLMs, broad plot premises are recovered more reliably than event-complete synopses; consistency across subtitle languages varies by model and language; national rating systems expose distinct calibration patterns; and strict evidence grounding varies strongly across language-safety categories.

Dataset summary

Property	Value
Films	1,012
Subtitle languages per film	6
Subtitle tracks	6,072
Timestamped subtitle entries	8,126,494
National rating systems	10
Split	test
License	CC BY-NC-SA 4.0

Languages

Every film contains one complete subtitle track for each of the following languages:

• English (en)
• Arabic (ar)
• Indonesian (id)
• Persian (fa)
• Romanian (ro)
• Vietnamese (vi)

Loading the dataset

from datasets import load_dataset

dataset = load_dataset("tafseer-nayeem/CineSubBench", split="test")
print(dataset)
print(dataset[0]["title"])

The repository also stores the canonical release as data/CineSubBench.json. Its SHA-256 checksum is recorded in checksums.sha256.

Data structure

Each row represents one film.

Film metadata

Field	Type	Description
id	Integer	Unique sequential CineSubBench film identifier.
title	String	Film title.
releaseYear	String	Film release year.
duration	String	Human-readable runtime.
imdbRating	String	IMDb user-rating value recorded during collection.
countriesOfOrigin	String	Production country or countries.
originalLanguages	String	Original spoken language or languages.
directors	Array of strings	Credited directors.
writers	Array of strings	Credited writers.
topCast	Array of objects	Principal cast entries with actor and character.
fullCast	Array of objects	Expanded cast entries with actor, characters, and additionalTexts.

Narrative and classification references

Field	Type	Description
plot	String	Concise premise-level narrative reference.
synopsis	String	Longer, spoiler-aware account of the narrative arc.
storyline	String	Short source-provided storyline description.
keyMessage	String	Reference statement of the principal theme or lesson.
genres	Array of strings	Gold multi-label genre set.
topics	String	Source-provided topical descriptors.
ageSuitabilityRating	String	General age threshold, such as 13+.
motionPictureRatings	Array of objects	Labels for ten national systems. Each item contains country, label, minimumAge, and ordinal.

Language-content assessment

Field	Type	Description
languageContentReference	Object	External review information retained for audit, containing rating, summary, and source.
languageContentAssessment	Object	Audited annotations derived from the English subtitle track.
inputLanguage	String	Language used to construct the annotations.
categories	Object	Annotations for strongProfanity, crudeBodilyLanguage, mildObscenity, and religiousProfanityAndExclamation.
occurrenceCount	Integer	Number of matched expressions in a category.
evidenceLineCount	Integer	Number of distinct subtitle entries containing category evidence.
evidenceSubtitleIndices	Array of integers	Indices linking evidence to English subtitle entries.

Sources and multilingual subtitles

Field	Type	Description
sources	Object	Per-film provenance URLs: imdbLink, commonSenseLink, and kidsInMindLink.
subtitles	Array of objects	Six subtitle tracks, each containing language, downloadCount, and entries.
language	String	Track language in Language name (code) form.
downloadCount	Integer	Provider download count retained as a subtitle-selection quality signal.
entries	Array of objects	Ordered subtitle units containing index, timeframe, and content.

Data sources

Source	Dataset fields
Common Sense Media	ageSuitabilityRating, storyline
Kids-in-Mind	Initial film inventory, keyMessage, languageContentReference, and language-content counts and glossary definitions used to audit the subtitle-derived languageContentAssessment
IMDb	Film metadata, narrative references, genre labels, national ratings, and cast and crew fields
OpenSubtitles and SubDL	Multilingual timestamped subtitle tracks

Each film retains source URLs for provenance and auditing.

Construction and quality assurance

CineSubBench integrates complementary information through IMDb title identifiers. Kids-in-Mind supplies the initial film inventory and language-content reference counts; Common Sense Media supplies age suitability and storylines; IMDb supplies film metadata, plot, synopsis, runtime, and regional ratings. OpenSubtitles is the primary subtitle source, with SubDL used for recovery.

Figure 2. Cross-source entity resolution, progressive filtering, complete-coverage optimization, subtitle verification, and task-oriented label curation produce the matched 1,012-film benchmark.

1. Resolve and filter films. Cross-source linking and task-metadata checks reduce an initial 6,186 films to a metadata-complete cohort of 2,322.
2. Optimize matched coverage. Complete-coverage searches select six subtitle languages (1,231 films) and ten national rating systems (1,012 films) over the same film instances.
3. Acquire and verify subtitles. Source-reported download counts prioritize candidates when several files exist. Structural, temporal, encoding, language-consistency, density, and runtime checks flag files for manual inspection, repair, cross-provider recovery, or exclusion.
4. Curate task references. Narrative fields are cleaned to retain film-internal content; country ratings are normalized within their own national systems. English language-safety evidence is derived from indexed subtitles using glossary-guided rules and reconciled with external counts through manual review of discrepancies.
5. Sanitize release text. The released subtitle tracks exclude 6,594 non-narrative distribution artifacts. In 65 otherwise narrative-relevant entries, an email address was redacted while the surrounding text was retained. Subtitle indices remain unchanged, preserving evidence links.

Intended uses

CineSubBench is intended for non-commercial research on long-form film understanding, multilingual subtitle processing, culturally situated audience assessment, and auditable language-safety evaluation. Its timestamped multilingual tracks can also support subtitle alignment, cross-lingual retrieval, temporal reasoning, summarization, and accessibility-oriented subtitle systems. Country-specific ratings retain their national label spaces, and language-safety evidence supports transparent human review.

Design scope

• Matched coverage. The benchmark prioritizes complete coverage across six subtitle languages and ten national rating systems for the same 1,012 films, supporting direct comparisons across tasks, languages, and cultural settings.
• Subtitle interface. Film-length subtitles provide a common, temporally ordered linguistic input for API-based and open-weight LLMs. Visual action, cinematography, music, gestures, and other non-speech signals belong to a complementary multimodal setting.

License

The dataset is available exclusively for non-commercial research under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License. See LICENSE for the full repository notice.

Users are responsible for complying with the license and the applicable terms associated with the original sources.