Datasetpublic

IMJONEZZ/star-wars-dataset

Star Wars dialogue dataset - annotation layers (snapshot 2026-09-21) One master cue table serving diarization, ASR and dialogue-LLM training: every subtitle cue part of 15 titles with a forced-aligned time span, a character label and provenance. No audio or video is included - the source media is copyrighted. Clip paths in the manifests resolve once you rebuild the clips from your own copies with the pipeline code (export_asr.py, export_diarization.py). The text layers contain… See the full description on the dataset page: https://huggingface.co/datasets/IMJONEZZ/star-wars-dataset.

sourceHugging Faceupdated 1d agoView on Hugging Face

2likes38downloads

Dataset Card

Star Wars dialogue dataset - annotation layers (snapshot 2026-09-21)

One master cue table serving diarization, ASR and dialogue-LLM training: every subtitle cue part of 15 titles with a forced-aligned time span, a character label and provenance. No audio or video is included - the source media is copyrighted. Clip paths in the manifests resolve once you rebuild the clips from your own copies with the pipeline code (export_asr.py, export_diarization.py). The text layers contain subtitle dialogue; keep that in mind before changing this repo's visibility.

Coverage

Title (`movie` column)	Cue parts	Named	Trusted (gold/silver)
A New Hope 4K77	1423	99.8%	99.8%
The Empire Strikes Back 4K80	1563	99.6%	99.6%
Return of the Jedi 4K83	1056	99.9%	99.9%
Rogue One	1597	99.9%	99.9%
Solo	2054	100.0%	100.0%
Revenge of the Sith	1238	99.7%	99.7%
Attack of the Clones	1188	99.1%	99.1%
The Rise of Skywalker	1532	100.0%	100.0%
The Phantom Menace	1429	99.9%	99.9%
The Last Jedi	1619	99.5%	99.5%
The Force Awakens	1510	97.7%	97.7%
Obi-Wan Kenobi	2292	91.1%	75.5%
The Mandalorian and Grogu	1536	71.0%	69.5%
The Mandalorian Season 1	2285	90.1%	79.6%
The Mandalorian Season 2	2219	95.9%	19.6% (audit in progress)

"Named" counts all cue parts including sound captions. Labels that have not passed a human audit carry origin ...-unaudited and quality review; train only on gold/silver. Still to come: The Mandalorian S3, Andor.

Layout

—cues/<title>.jsonl - the master table: span, text, asrtext, speaker, origin, score, timing method, verbatim, ctcscore, quality. Everything else derives from it.
—cue_timing/ - wav2vec2 CTC forced alignment per cue part (with coverage).
—cue_labels/ - screenplay-match and SDH-tag speaker labels. knowledge_labels/ - LLM attribution from scene context (tiers high/medium/low) + AUDITED.txt (which titles/tiers passed the human audit). human_labels/ - reviewer verdicts (top precedence). speakers.json - canonical names and aliases.
—asr/ - train/val/test manifests (audio, text, duration, speaker, movie), separate held-out *_tv splits.
—diarization/ - pyannote protocol: RTTM, UEM, lists, database.yml (interleaved 10-minute dev blocks).
—script/ - screenplay-style corpus CHARACTER: line with GLM special tokens, script.jsonl, manifest.json.

Files are named by title with underscores (cues/The_Phantom_Menace.jsonl); the movie column holds the readable title. The original trilogy rows come from the Project 4K77 / 4K80 / 4K83 restorations. The viewer configs are cues (default), asr and script.

How labels were made

Speaker identity comes from text and context, never from voice clustering: screenplay-to-cue text matching (~95% precise), SDH caption tags, LLM attribution from scene context (high tier audited at 96-100%), then human review. Timing is forced alignment of the known cue text, because subtitle display times run late and drift.

Models trained from this data

IMJONEZZ/star-wars-qwen3-asr-1.7b, IMJONEZZ/star-wars-wespeaker-resnet34, IMJONEZZ/star-wars-pyannote-segmentation-3.0, IMJONEZZ/star-wars-z-image-turbo-lora.