NewsShorts-Bench

Test sets for each stage of the NewsShorts research program, which turns long news broadcasts into faithful short clips. Each stage of the pipeline is a measured study; this repo holds the data those studies use. Results, code and process notes are in NewsShorts-Bench-Results.

Status: growing. Sets are added as stages finish. Cite a specific commit.

What's here

Folder	Stage	What it is	Truth	Size
s2_asr/	S2, speech recognition	Reference windows (15-30 s) from human-written subtitles, multilingual	human subtitles	358 windows
s4_news_speakers/	S4, diarization	Speaker turns from labelled human subtitles of White House videos ("The President:", "Secretary Azar:")	human subtitles, partial	16 videos
s5_namebars/	S5, on-screen text	Synthetic news name bars (name + role) in 13 languages over real broadcast frames	exact (drawn by us)	520 images
s6_splitscreen/	S6, active speaker	Recipe for split-screen pairs: two people talking, one voice	exact (we chose the audio)	116 pairs
s7_synthetic/	S7, story segmentation	Synthetic broadcasts joined from single-story videos	exact	50 broadcasts
s7_real_silver/	S7, story segmentation	Story starts in real multi-story bulletins (English, Spanish, Russian)	model-made (silver)	21 bulletins, 7 h
s9_highlights_silver/	S9, highlight scoring	Per bulletin, stories ranked as standalone vertical shorts, with a 20-60 s clip for the top 3	model-made (silver)	19 bulletins, 174 stories
s10_traps/	S10/S11, faithfulness and repair	Per real story, a faithful clip, a shorter fair trim, and a misleading clip made by moving one edge (dropped correction, outcome, qualifier, denial, attribution)	model-made by construction, not human-checked	54 stories, 162 clips
s8_roles_silver/	S8, unit roles	Timelines of roles (anchor, package, soundbite, standup, interview, graphic, titles, other) over the same bulletins	model-made (silver)	21 bulletins, 7 h

Truth levels

• Exact: known by construction (S5, S6, S7 synthetic).
• Human, partial: written by people, but not for this task (S2, S4).
• Silver: made by a model and not checked by a person (S7 real, S8, S9): Gemini 4 Argon watching each bulletin with sound, under two differently worded prompts. Both answers are included, so disagreement is visible. Treat as provisional.

Details by set

s2_asr/asr_ref_windows.parquet: title, split, lang, window_id, start_s, end_s, ref_text. title is the Wikimedia Commons file; audio is not included.

s4_news_speakers/turns.jsonl: per video, commons_title, duration_s and turns (start, end, speaker). A label holds until the next one. "The Press" and "Audience Member" are mixed groups. Subtitle timing is coarse.

s5_namebars/: images/<id>.png is the bottom 40% of a 1280-wide frame (640 at 360p) with a drawn name bar; metadata.jsonl has file_name, lang, script, name, role, style, res, crf, size, upper, font. Backgrounds are blurred (the whole frame, and detected faces more), so no real person appears under an invented name. That also makes background text unreadable, so this variant is easier than the set the study scored. Reference scores on this exact variant (reference_scores.json): Cloud Vision 99.2% exact names (95% CI 98.5-99.8), PP-OCRv6 with the language given 89.4% (86.7-91.9). On the unblurred study set they were 97.3% and 85.8%. Background sources and licenses: backgrounds_attribution.json (54 public domain, 6 CC BY 3.0).

s6_splitscreen/: clips.json gives each 10 s source window (Commons URL, start time, labelled speaker); pairs.json gives each pair (left, right, crop offsets left_x/right_x, and audio, the side whose sound is kept). The videos aren't redistributed; eval/asd_split.py in the results repo rebuilds them from Commons.

s7_synthetic/broadcasts.json: each broadcast's stories in order, with boundaries as subtitle-line indices and in seconds. subtitle_pages.json lists the Commons subtitle pages; eval/story_seg.py build recreates the text.

s8_roles_silver/timelines.jsonl: per bulletin, consecutive segments (start, end in seconds, role) under two prompts (segments_a, segments_b); role definitions are in eval/unit_roles.py in the results repo.

s9_highlights_silver/rankings.jsonl: per bulletin, the stories (S7 silver starts, titles and sign-offs dropped by S8 roles) and two prompts' answers (result_a, result_b): a full ranking and clip spans (MM:SS) for the top 3. Model preference, not audience data.

s10_traps/: stories.jsonl has each story's sentences (start, end, text) from Chirp 3 word times, with the Commons title and license; items.jsonl has three clips per story (kind: faithful, trimmed, misleading; label: true for misleading), as inclusive sentence indices first/last, the clip text, and for misleading clips the trap_type and the false_impression it creates. Built by Gemini 4 Argon under fixed rules checked in code (eval/faithfulness.py); a second Argon prompt agreed on all three clips for 26 of the 54 stories, so treat labels as provisional. Every word of a clip is in the story: that is the point (grounding checkers score them all as supported).

s7_real_silver/labels.jsonl: per bulletin, the Commons title and URL, language, duration, license, and story starts with headlines under both prompts (stories_a, stories_b).

Licensing

CC BY-SA 4.0. Source material keeps its own terms: VOA and White House videos are US government works in the public domain; some sources are CC BY 3.0, and Commons subtitles are CC BY-SA. Attribution is listed per file where it applies. Names and roles in s5_namebars are invented (Faker) and refer to no real person.

Limits

• Small sets, for comparing methods, not for training.
• S7 real, S8, S9 and S10 labels are model-made; S4 turns use coarse subtitle timing.
• Mostly English outside S2 and S5.