Omni-Sets

A large-scale, multi-modal instruction-tuning dataset spanning six modalities (audio, speech, image, video, visual documents, and cross-modal omni) with both single-turn dense captions and multi-turn instruction-following conversations. Designed for training omni-modal language models that can perceive and reason across all modalities.

590,858 total samples | 5,635 hours of audio/video | 6 configs | 17 source datasets

Overview

Config	Modality	Samples	Caption	Chat (IT)	Media Duration
audio	Audio	135,587	45,193 (33%)	90,394 (67%)	622.6 hours
speech	Speech	140,219	46,740 (33%)	93,479 (67%)	339.7 hours
image	Image	157,782	86,143 (55%)	71,639 (45%)	--
video	Video	65,611	35,693 (54%)	29,918 (46%)	1,835.9 hours
omni	Video+Audio+Speech	40,087	--	40,087 (100%)	2,836.8 hours
visual_doc	Document Image	51,572	28,311 (55%)	23,261 (45%)	--
Total	All	590,858	242,080	348,778	5,635.0 hours

---

Generation Pipeline

All samples are generated through a two-stage pipeline:

1. Stage 1, Dense Captioning: Raw media (audio/image/video) is sent to Qwen3-Omni (multimodal) along with the original ground-truth caption for hallucination prevention. The model produces semantically rich, grounded descriptions.

2. Stage 2, Instruction Tuning: Dense captions are fed to Qwen3-Instruct (text-only) to generate natural multi-turn conversations in 8+ stylistic variants (casual, analytical, negation/correction, practical, accessibility, comparison, scene reconstruction, quick Q&A).

The omni config uses a different pipeline: video + extracted audio are sent simultaneously to Qwen3-Omni, which generates cross-modal multi-turn Q&A requiring both visual and auditory understanding. 24.6% of omni samples additionally have TTS-synthesized speech questions (Kokoro, 19 voices).

This ensures captions are grounded in actual media content (not text-only hallucinations) while IT conversations are diverse and natural.

---

Conversation Statistics

Turn Distribution

User Turns	Audio	Speech	Image	Video	Omni	Visual Doc
1 (caption)	45,393 (33.5%)	46,826 (33.4%)	86,143 (54.6%)	35,693 (54.4%)	3,034 (7.6%)	28,311 (54.9%)
2-3	12,631 (9.3%)	27,165 (19.4%)	23,729 (15.0%)	13,884 (21.2%)	21,018 (52.4%)	11,287 (21.9%)
4+	77,563 (57.2%)	66,228 (47.2%)	47,910 (30.4%)	16,034 (24.4%)	16,035 (40.0%)	11,974 (23.2%)

• Single-turn (caption): User asks a brief question, assistant provides a dense description of the media.
• Multi-turn (chat/IT): Natural back-and-forth conversations with follow-ups, corrections, comparisons, and practical assessments. Most common length is 3-4 user turns (6-8 messages total).

IT Prompt Variants

Conversations are generated from 8 distinct prompt styles per modality to maximize diversity:

Style	Description	Example User Turn
Casual	Open-ended curiosity	"What's that noise?" / "Cool photo, what's going on?"
Negation	User guesses wrong, assistant corrects	"Is that a cat?" (it's a dog)
Analytical	Detailed breakdowns and counting	"How many people are there?"
Practical	Task-oriented assessment	"Could I use this as a thumbnail?"
Scene	Setting and context inference	"Where do you think this was recorded?"
Quick	2-3 short exchanges	"Indoor or outdoor?"
Comparison	Contrasting elements within media	"How does the beginning differ from the end?"
Accessibility	Rich descriptions for captioning	"Describe this for someone who can't see it"

Speech-specific additional variants: transcription, speaker analysis, comprehension, quality/environment.

Omni-specific prompt variants: visual-audio synchronization, speaker-visual relationships, narrative arc combining audio & visuals, deep cross-modal comprehension.

---

Source Datasets

Sample counts below reflect the slice used in Omni-Sets after filtering, deduplication, and removal of corrupted media. They do not necessarily match the upstream dataset's full size. The License column reflects the license declared by the original creator (verified against the upstream paper/repo, not just the HuggingFace mirror); entries are blank where no clear license is declared upstream. Individual source datasets retain their respective licenses regardless of the Omni-Sets aggregate license.

Audio (135,587 samples, 622.6 hours)

Source	HF Path	Samples	License	Description
WavCaps (FreeSound)	cvssp/WavCaps	78,642	Academic / research-only	~262K Freesound clips re-captioned via ChatGPT pipeline. Diverse environmental sounds: nature, urban, household, industrial. Sub-collection licenses (CC0/CC-BY/CC-BY-NC) vary by clip.
WavCaps (AudioSet-SL)	cvssp/WavCaps	32,493	Academic / research-only	Strongly-labeled subset of AudioSet (~108K clips in upstream) re-captioned. Music, speech, animal, vehicle, and mechanical sounds with verified event annotations.
AudioCaps	OpenSound/AudioCaps	13,553	MIT (academic only)	Human-written captions over AudioSet clips (Kim et al., NAACL 2019). Covers everyday sounds: alarms, footsteps, weather, appliances, wildlife.
WavCaps (BBC)	cvssp/WavCaps	9,355	Academic / research-only	BBC Sound Effects (~31K clips upstream), re-captioned. Broadcast-quality recordings of nature, transport, crowds, weather, and foley.
Clotho	CLAPv2/clotho_full_2	1,182	Tampere University license (audio: mixed CC from Freesound)	Clotho v2 (Drossos et al., ICASSP 2020): 6,974 Freesound clips, 15-30s each, 5 human-written captions per clip. Captions released under the Tampere University Dataset License; underlying audio inherits its original Freesound CC license per clip.
WavCaps (SoundBible)	cvssp/WavCaps	362	Academic / research-only	SoundBible clips (~1.2K upstream), re-captioned. Royalty-free sound effects: bells, clicks, splashes, individual events.

Duration stats: mean 16.5s, median 10.0s (P25: 10.0s, P75: 30.0s, max: 60.0s)

Diversity rationale: WavCaps provides scale and variety across 4 sub-collections (environmental, professional, crowd-sourced). AudioCaps adds human-written grounding. Clotho offers high-quality curation. Together they cover the full spectrum of non-speech, non-music audio.

Speech (140,219 samples, 339.7 hours)

Source	HF Path	Samples	License	Description
TED-LIUM	sanchit-gandhi/tedlium-data	80,479	CC BY-NC-ND 3.0	TED-LIUM 3 (Hernandez et al., 2018): 452h of TED talk segments. Diverse speakers, accents, and topics (science, technology, culture, politics). Professional recording quality.
HeySQuAD	yijingwu/HeySQuAD_human	21,596	CC BY 4.0 (HF declared)	HeySQuAD (Wu et al., 2023): 76K human-spoken SQuAD questions read by 12 native English speakers. Conversational, question-oriented speech. Underlying SQuAD is CC BY-SA 4.0.
Spoken Alpaca	GSQA/spoken-alpaca-gpt4	15,404	CC BY-NC 4.0	TTS reading of vicgalle/alpaca-gpt4 (52K GPT-4 instruction-response pairs, Peng et al., 2023). Technical, creative, and analytical content in clear speech.
Gemini Speech	shb777/gemini-flash-2.0-speech	14,178	Apache-2.0	Synthetic speech generated by Gemini 2.0 Flash (Kore/Puck voices). Clean, diverse content including code explanations, stories, and instructions.
LibriSpeech	openslr/librispeech_asr	8,562	CC BY 4.0	LibriSpeech (Panayotov et al., 2015), train-clean-100 subset. Read English audiobooks (LibriVox), clear articulation, varied speakers. Gold standard for ASR.

Duration stats: mean 8.7s, median 6.5s (P25: 4.4s, P75: 9.6s, max: 60.0s)

Diversity rationale: TED-LIUM provides scale with natural, spontaneous speech across topics and accents. HeySQuAD adds question-form speech patterns. Spoken Alpaca and Gemini cover instruction-following content. LibriSpeech anchors with clean, well-articulated read speech. The mix spans spontaneous-to-read, question-to-monologue, and human-to-synthetic speech styles.

Image (157,782 samples)

Source	HF Path	Samples	License	Description
ShareGPT-4o Images	OpenGVLab/ShareGPT-4o	77,539	MIT (academic only)	OpenGVLab caption-quality data generated by prompting GPT-4o on diverse images. Real-world content including photos, documents, charts, screenshots, and UI elements.
Recap-COCO-30K	UCSC-VLAA/Recap-COCO-30K	41,195	CC BY 4.0	COCO 2014 val (30,500 images) with LLaMA-3 generated dense recaptions (Li et al., "What If We Recaption Billions of Web Images with LLaMA-3?"). Everyday scenes: people, animals, food, vehicles, sports.
Flickr30K	clip-benchmark/wds_flickr30k	39,048	Research-only (Flickr ToU)	Flickr30K (Young et al., TACL 2014): 31,783 Flickr photos with 5 human captions each (158,915 captions total). People-centric: activities, events, occupations, outdoor scenes.

Diversity rationale: ShareGPT-4o covers the long tail of image types (screenshots, charts, memes, documents) that natural photo datasets miss. COCO provides the standard distribution of everyday scenes and objects. Flickr30K adds people-centric, activity-focused content. Together they span photos, documents, charts, UI, and creative content.

Video (65,611 samples, 1,835.9 hours)

Source	HF Path	Samples	License	Description
ActivityNet Captions	friedrichor/ActivityNet_Captions	50,577	No explicit license stated (devkit is MIT); videos follow YouTube ToS	ActivityNet Captions (Krishna et al., ICCV 2017): 20K YouTube videos / 849h with dense temporal segment captions (~100K segments total). Sports, cooking, grooming, music, crafts, daily activities. The authors publish no license for the annotations; the ActivityNet devkit code is MIT. The underlying video clips are subject to YouTube's Terms of Service.
MSR-VTT	friedrichor/MSR-VTT	12,176	Research-only (Microsoft Research License Agreement)	MSR-VTT (Xu et al., CVPR 2016): 10K web video clips, 10-20s each, 20 captions per clip (200K total). 20 categories: music, news, sports, cooking, gaming, science, travel. MSR-LA permits non-commercial use only and requires Microsoft's prior written authorization to redistribute.
ShareGPT-4o Videos	OpenGVLab/ShareGPT-4o	2,858	MIT (academic only)	Caption-quality video data from prompting GPT-4o. Tutorials, vlogs, screen recordings, presentations.

Duration stats: mean 100.7s, median 92.3s (P25: 26.0s, P75: 165.0s, max: 755.3s)

Diversity rationale: ActivityNet provides temporal grounding with segment-level annotations across 200+ activity categories. MSR-VTT adds breadth with 20 diverse categories in short-form clips. ShareGPT-4o contributes screen recordings and presentation-style content that activity datasets lack.

Omni (40,087 samples, 2,836.8 hours)

Source	HF Path	Samples	License	Description
FineVideo	HuggingFaceFV/finevideo	40,087	Mixed Creative Commons (gated)	FineVideo (Hugging Face FV): 43,751 Creative-Commons YouTube videos with rich audio. Diverse categories: entertainment, education, science, news, lifestyle, sports, art, automotive. Long-form (mean ~4.2 min).

Duration stats: mean 254.8s, min 14.0s, max 600.0s

Speech questions: ~25% of samples have TTS-synthesized spoken questions via Kokoro (19 American English voices: 10 female, 9 male). The remaining ~75% use text questions.

Categories: Entertainment (8,394), Education (6,186), Science & Technology (4,958), News & Politics (4,537), Lifestyle (4,350), Hobbies & Interests (4,219), Sports (3,134), Art & Creativity (2,967), Automotive (1,342).

Cross-modal design: Every Q&A pair requires both visual AND auditory understanding; answering from only video or only audio is incomplete. Five prompt variants rotate through visual-audio synchronization, speaker-visual relationships, narrative arc, deep cross-modal comprehension, and quick Q&A.

Visual Document (51,572 samples)

Source	HF Path	Samples	License	Description
ColPali Train	vidore/colpali_train_set	22,720	Mixed (per sub-source)	ColPali training set (Faysse et al., 2024, arXiv:2407.01449): ~119K query-image pairs across 5 sub-sources: DocVQA, InfoVQA, TATDQA, ArXivQA, plus synthetic web-PDF QA generated with Claude-3 Sonnet.
VisRAG Train	openbmb/VisRAG-Ret-Train-In-domain-data	28,852	Mixed (per sub-source)	VisRAG-Ret in-domain training data (Yu et al., 2024): 122,752 query-document pairs across 6 sub-sources: PlotQA, ArXivQA, InfoVQA, MP-DocVQA, SlideVQA, ChartQA.

Design: 43,444 unique document images split into three roles:

• Caption-only (20,183 images): Dense structured description of the document (layout, text, figures, tables, styling)
• Instruction-only (15,133 images): Multi-turn chat grounded in the document content
• Bridge (8,128 images): Both caption AND instruction. The same image appears in both, bridging understanding of dense description and conversational Q&A

Sub-sources: arxiv_qa, docvqa, pdf, tatdqa, Infographic-VQA (ColPali) | ArxivQA, ChartQA, InfoVQA, MP-DocVQA, PlotQA, SlideVQA (VisRAG)

---

Schema

audio / speech

Column	Type	Description
id	string	Unique sample identifier ({source}_{idx}_{hash})
media	Audio	Audio waveform (16kHz mono WAV)
messages	list[{role, content}]	Conversation (caption or multi-turn IT)
modality	string	audio or speech
category	string	caption (single-turn) or chat (multi-turn IT)
original_text	string	Ground-truth caption from source dataset
source_dataset	string	Source dataset name
source_hf_path	string	HuggingFace dataset path
duration_s	float	Audio duration in seconds
sampling_rate	int	Always 16000

image / video

Column	Type	Description
id	string	Unique sample identifier
media	Image/Video	JPEG image or MP4 video (h264)
messages	list[{role, content}]	Conversation (caption or multi-turn IT)
modality	string	image or video
category	string	caption (single-turn) or chat (multi-turn IT)
original_text	string	Ground-truth caption from source dataset
source_dataset	string	Source dataset name
source_hf_path	string	HuggingFace dataset path
duration_s	float	Video duration in seconds (0 for images)
fps	float	Video frame rate (0 for images)

omni

Column	Type	Description
id	string	Unique sample identifier (finevideo_{youtube_id})
media	Video	MP4 video with audio track (h264)
messages	list[{role, content}]	Multi-turn cross-modal Q&A
modality	string	Always omni
category	string	Content category (Entertainment, Education, etc.)
original_text	string	Video description from FineVideo metadata
source_dataset	string	Always finevideo
source_hf_path	string	HuggingFaceFV/finevideo
duration_s	float	Video duration in seconds
fps	float	Video frame rate
has_speech_question	bool	Whether first user question is spoken audio
question_audio	Audio	TTS-synthesized spoken question (16kHz WAV), null if text-only
question_transcript	string	Text transcript of the spoken question
speaker	string	Kokoro TTS voice name (e.g., am_adam, af_bella)

visual_doc

Column	Type	Description
id	string	Unique sample identifier
media	Image	Document image (JPEG/PNG)
messages	list[{role, content}]	Conversation (caption or multi-turn IT)
modality	string	Always visual_doc
category	string	caption (dense description) or chat (multi-turn IT)
source_dataset	string	colpali or visrag
source_hf_path	string	HuggingFace dataset path
source_subset	string	Specific sub-source (e.g., arxiv_qa, ChartQA)

---

Usage

from datasets import load_dataset

# Load individual configs
audio = load_dataset("MBZUAI/Omni-Sets", name="audio", split="train")
speech = load_dataset("MBZUAI/Omni-Sets", name="speech", split="train")
images = load_dataset("MBZUAI/Omni-Sets", name="image", split="train")
videos = load_dataset("MBZUAI/Omni-Sets", name="video", split="train")
omni = load_dataset("MBZUAI/Omni-Sets", name="omni", split="train")
visual_doc = load_dataset("MBZUAI/Omni-Sets", name="visual_doc", split="train")

# Filter by category
captions_only = audio.filter(lambda x: x["category"] == "caption")
chat_only = audio.filter(lambda x: x["category"] == "chat")

# Filter by source
ted = speech.filter(lambda x: x["source_dataset"] == "tedlium")

# Omni: speech-question samples only
speech_qs = omni.filter(lambda x: x["has_speech_question"])

License

The annotations we generated — the dense captions and multi-turn conversations, produced with Qwen3-Omni and Qwen3-Instruct (both Apache-2.0, outputs unrestricted) — are released under CC BY-SA 4.0.

The underlying media is not ours to relicense. Omni-Sets is an aggregation assembled for research convenience: it bundles media from 17 source datasets that were otherwise scattered across different hosts, formats and access paths. Each source retains its own license and terms, listed per-row in the Source Datasets tables above, and those terms govern the media regardless of the license on this aggregate. Several are materially more restrictive than CC BY-SA 4.0.

If you use Omni-Sets, you are responsible for complying with the terms of each source dataset you actually use. For commercial use in particular, do not rely on the aggregate license — check the per-source rows and filter accordingly (source_dataset and source_hf_path are present on every sample for exactly this purpose). If you are a rights holder and want material removed, please open an issue on this repository and we will take it down.

Citation

If you use this dataset, please cite:

@inproceedings{omniembedmini2026,
  title     = {Omni-Embed-Mini: Binding Modalities Without Forgetting via Dense Distillation},
  author    = {TBD},
  booktitle = {TBD},
  year      = {2026},
  note      = {Camera-ready in preparation}
}