SparkX2.6-MM

Dataset-only training collection for the English/Chinese multimodal adaptation of Spark-X2.6-4B. Images/spatial grounding, real video/time lapse, speech, environmental sound, music and native waveform targets are stored as separately attributed components. Only components in the table below are present in this revision; ongoing acquisitions are not counted as released data.

This corpus is prepared for the initial full-parameter campaign. Dataset preparation and packaging checks do not certify model quality, native decoder integration, fresh-VM optimizer recovery or successful RSI. Training scripts, notebooks, model weights and runtime bundles are excluded from the current dataset tree.

Released components

Component	Physical source packets	Instruction variants	TAR payload GB	Coverage and limits
DOCCI detailed captions	8,776	8,776	2.104	8,776 original human captions/images; strict local and upstream training intersection; historical exposure retained
Open Images / Localized Narratives	100,000	100,000	31.644	Original human final transcripts; per-image photographer/title/license/source retained; CVDF JPEGs retained; known COCO-CN evaluation aliases and existing-core photos excluded
COCO / COCO-CN and controlled geometry	7,192	42,215	1.182	Human EN/ZH captions and boxes; complete-instance-derived counts; renderer-truth points/boxes/motion/occlusion/configured elapsed time; original COCO-CN val/test identities excluded
SA-V / PLM training video	919	2,358	7.562	919 real source-training clips; manual-mask box tracks, human region captions and temporal localization; no audio; source and actual presentation clocks preserved
CREMA-D synchronized audiovisual affect	2,100	2,100	0.590	2,100 original acted English clips from 70 selected actors; synchronized waveform/video and original crowd labels; other actors excluded; not broad AV event QA
LibriTTS-R / AISHELL-3 bilingual speech	32,890	65,780	4.738	16,020 EN +16,870 Mandarin real waveform/transcript pairs; 25.277h EN +16.998h ZH; ASR and speech generation share each waveform; lossless FLAC, original sample rates; LibriTTS-R includes upstream restoration
FSD50K environmental sound	9,911	19,822	6.939	Original upstream training clips; 200 positive sound-event classes; understanding and waveform-generation variants; human weak clip labels, not human free-form captions; silent, duplicate and numerically flagged waveforms excluded by declared RMS/peak/PCM-extrema gates
MusicNet / open-licensed music	1,027	4,108	0.731	1,027 real 15-second excerpts (4.279h) from 100 classical performances, 52 work groups, seven composers and nine active instruments; 271 excerpts have separately evidenced performer CC0 grants, 756 retain publisher CC BY4; score-derived scripted EN/ZH understanding/generation tasks; official test works and sibling recordings excluded; limited genre coverage
NASA/JSC real Earth-observation time lapse	72	360	0.466	Real source video derivatives with original full-sequence NASA descriptions/context; reviewed derivatives omit title/insignia slates; contextual GMT arithmetic/calibration tasks do not establish visually inferred elapsed time; source capture spans are not per-frame derivative wall-clock labels; narrow Earth-observation domain
Bilingual document OCR / region reading	9,671	19,342	2.261	9,671 synthetic rendered EN/ZH documents, not natural scans; 19,342 OCR/region-reading instructions; 329 missing-glyph candidates excluded; source articles grouped with text replay
Human EN/ZH text replay	20,000 source articles	40,000 passages	JSONL gzip	20,000 EN +20,000 ZH original human continuations; shares article groups with rendered OCR; not human instruction annotation

Counting: an ASR/TTS pair is two instructions over one recording. A video with several questions remains one source packet. Some packets contain several frame/audio/video assets. These counts are not independent real-world event counts or model-consumed record counts. Source groups remain conservative, and pretraining overlap/unknown crops cannot be ruled out.

Format and source roles

Each uncompressed USTAR shard is at most 1 GiB and contains adjacent media plus JSON. RELEASE.json binds exact shard sizes/SHA256, every member/byte offset, source manifests, license evidence and samples.jsonl. Use the complete actual repository commit and the index/catalog/shard hashes; moving main is not a frozen training identity.

Grouped packets use grouped_schema=spark_shared_media_task_variants_v1. task_variants holds every original trainable instruction, its media order, clocks, language, targets and input/output roles. Flatten or sample these variants while sharing the physical media. The top-level prompt is a compatibility default and is not an additional instruction. Forward/reverse frame order is preserved independently per variant.

For native generation, target.kind=waveform, target.path, audio_direction=text_to_audio and media_role=target identify the target waveform. Do not feed target audio into conditioning. Recognition/understanding variants use the waveform as input. No pretrained codec latents or frozen vision features are substituted for source media; codecs/processors belong to the separately qualified trainer.

Keep active TARs, indexes and decoded working buffers on Colab local disk; Google Drive is cold storage for closed artifacts/checkpoints. The 230-GB runtime requires bounded rolling staging, not a full corpus clone. Shard arrival/partials and two complete measured checkpoints must preserve the 44-GB free floor.

Source citations, licenses and transformations

Licenses are component- and asset-specific. This collection does not relicense every source under one blanket license. Per-asset creator/title/source/license, modifications, exact media hashes and local source-evidence hashes are in RIGHTS_LEDGER.json and the JSON records. Original grants/notices are retained under each component; source manifests identify pinned mirrors separately from original publishers.

• DOCCI detailed captions: Google LLC; Onoe et al. (2024), DOCCI: Descriptions of Connected and Contrasting Images. Original source, paper or additional primary documentation. Terms: CC BY 4.0. 8,776 original human captions/images; strict local and upstream training intersection; historical exposure retained.
• Open Images / Localized Narratives: Pont-Tuset et al. (2020), Connecting Vision and Language with Localized Narratives; Kuznetsova et al. (2020), Open Images. Original source, paper or additional primary documentation. Terms: Images: individual CC BY 2.0; annotations: CC BY 4.0. Original human final transcripts; per-image photographer/title/license/source retained; CVDF JPEGs retained; known COCO-CN evaluation aliases and existing-core photos excluded.
• COCO / COCO-CN and controlled geometry: Lin et al. (2014), Microsoft COCO; Li et al. (2019), COCO-CN for Cross-Lingual Image Tagging, Captioning and Retrieval. Original source, paper or additional primary documentation. Terms: Photos: CC BY 2.0; COCO annotations: CC BY 4.0; COCO-CN: MIT; controlled media: CC0. Human EN/ZH captions and boxes; complete-instance-derived counts; renderer-truth points/boxes/motion/occlusion/configured elapsed time; original COCO-CN val/test identities excluded.
• SA-V / PLM training video: Ravi et al. (2024), SAM 2; Meta PLM-Video-Human original training annotations. Original source, paper or additional primary documentation. Terms: CC BY 4.0. 919 real source-training clips; manual-mask box tracks, human region captions and temporal localization; no audio; source and actual presentation clocks preserved.
• CREMA-D synchronized audiovisual affect: Cao et al. (2014), CREMA-D: Crowd-Sourced Emotional Multimodal Actors Dataset. Original source, paper or additional primary documentation. Terms: Database: ODbL 1.0; individual content: DbCL 1.0. 2,100 original acted English clips from 70 selected actors; synchronized waveform/video and original crowd labels; other actors excluded; not broad AV event QA.
• LibriTTS-R / AISHELL-3 bilingual speech: Koizumi et al. (2023), LibriTTS-R; Shi et al. (2020), AISHELL-3. Original source, paper or additional primary documentation. Terms: LibriTTS-R: CC BY 4.0; AISHELL-3: Apache 2.0. 16,020 EN +16,870 Mandarin real waveform/transcript pairs; 25.277h EN +16.998h ZH; ASR and speech generation share each waveform; lossless FLAC, original sample rates; LibriTTS-R includes upstream restoration.
• FSD50K environmental sound: Fonseca et al. (2022), FSD50K: An Open Dataset of Human-Labeled Sound Events. Original source, paper or additional primary documentation. Terms: Per recording: CC0 or CC BY 3.0; annotation terms retained. Original upstream training clips; 200 positive sound-event classes; understanding and waveform-generation variants; human weak clip labels, not human free-form captions; silent, duplicate and numerically flagged waveforms excluded by declared RMS/peak/PCM-extrema gates.
• MusicNet / open-licensed music: Thickstun et al. (2017), Learning Features of Music from Scratch; original MusicNet publishers and credited performers. Original source, paper or additional primary documentation. Terms: MusicNet publisher CC BY 4.0; separately identified CC0 performances where supported. 1,027 real 15-second excerpts (4.279h) from 100 classical performances, 52 work groups, seven composers and nine active instruments; 271 excerpts have separately evidenced performer CC0 grants, 756 retain publisher CC BY4; score-derived scripted EN/ZH understanding/generation tasks; official test works and sibling recordings excluded; limited genre coverage.
• NASA/JSC real Earth-observation time lapse: NASA / Johnson Space Center, Crew Earth Observations Videos; each original clip credited individually. Original source, paper or additional primary documentation. Terms: NASA media-use policy / applicable US Government material; third-party exceptions retained. Real source video derivatives with original full-sequence NASA descriptions/context; reviewed derivatives omit title/insignia slates; contextual GMT arithmetic/calibration tasks do not establish visually inferred elapsed time; source capture spans are not per-frame derivative wall-clock labels; narrow Earth-observation domain.
• Bilingual document OCR / region reading: Wikipedia contributors (2023), original English/Chinese text; Noto CJK fonts; deterministic project document renderings. Original source, paper or additional primary documentation. Terms: CC BY-SA 3.0 per pinned source card; font SIL OFL 1.1 notice retained. 9,671 synthetic rendered EN/ZH documents, not natural scans; 19,342 OCR/region-reading instructions; 329 missing-glyph candidates excluded; source articles grouped with text replay.

Direct performer evidence for the CC0 music subset: Kimiko Ishizaka Open Well-Tempered Clavier release and the performer's album page. The recording grant and MusicNet's aligned-note annotation grant remain separate; unchanged original crops were verified sample by sample.

The FSD50K selection retains 9,911 recordings (21.6524h): 4,340 CC0 and 5,571 CC BY3, covering all 200 classes across 1,883 uploader families. After silent/duplicate exclusions, 59 recordings with more than 1% PCM-extrema samples and 27 with both RMS below -60dBFS and peak below -40dBFS were excluded. These are measured numerical gates, not a human listening or general fidelity certification. Original WAV bytes and English weak labels remain unchanged; no Chinese free-form sound captions are claimed. The original FSD50K notice asking commercial users to contact its authors is retained.

• Wikipedia text: Wikipedia contributors, 20231101 English/Chinese dumps, wikimedia/wikipedia. Article URLs/titles and original selected text are retained. Source card: CC BY-SA3.0/GFDL. Text is excerpted into exact continuations; document derivatives normalize whitespace and render new layouts. Known English retention and WikiText validation/test article identities were checked separately; this is not a claim of absence of model pretraining overlap.

The CREMA-D database/content component keeps ODbL/DbCL terms separately. Wikipedia-derived passages and page renderings keep their ShareAlike notices. CC0 sound and separately identified CC0 musical performances are not the same license category as CC BY recordings. A public-domain composition does not establish that a particular performance recording is public domain. NASA media remain subject to the applicable NASA reuse policy, credits and stated third-party exceptions.

Split, quality and capability boundaries

Only admitted source-training examples are included. Protected project evaluation targets are excluded. Original COCO-CN evaluation photo identities and linked groups are excluded across tasks and known Open Images aliases. Public evaluation recordings and sibling musical-work identities are excluded where documented. All variants/transforms retain source-family groups; source articles are shared across OCR/text and must not be split independently.

Labels retain their origin: original human captions/transcripts, human weak sound labels, score-derived musical labels, original crowd acted-affect annotations, annotation-derived counts/boxes, or explicitly synthetic renderer truth. Template translations are identified as templates rather than human bilingual captions. Playback timestamps are distinct from capture time; unknown capture clocks remain unknown. Weak audio labels are not dense event boundaries, and actual waveform targets do not themselves prove generation quality.

Instructor/RSI corrections depend on a trained parent and independent evaluation. They are not fabricated in advance or mislabeled as human ground truth in this initial source corpus. The reserved evaluation data, teacher key, old workstation checkpoints and runtime code are not included.

Bibliography

Cite the exact dataset commit and originating sources. Component bibliography entries are reproduced below for convenience.

@inproceedings{OnoeDocci2024,
  author        = {Yasumasa Onoe and Sunayana Rane and Zachary Berger and Yonatan Bitton and Jaemin Cho and Roopal Garg and
    Alexander Ku and Zarana Parekh and Jordi Pont-Tuset and Garrett Tanzer and Su Wang and Jason Baldridge},
  title         = {{DOCCI: Descriptions of Connected and Contrasting Images}},
  booktitle     = {ECCV},
  year          = {2024}
}

@inproceedings{ponttuset2020connecting, title={Connecting Vision and Language with Localized Narratives}, author={Pont-Tuset, Jordi and Uijlings, Jasper and Changpinyo, Soravit and Soricut, Radu and Ferrari, Vittorio}, booktitle={ECCV}, year={2020}, url={https://arxiv.org/abs/1912.03098}}
@article{kuznetsova2020openimages, title={The Open Images Dataset V4}, author={Kuznetsova, Alina and others}, journal={IJCV}, year={2020}, url={https://arxiv.org/abs/1811.00982}}

@misc{lin2014coco,
  title={Microsoft COCO: Common Objects in Context},
  author={Tsung-Yi Lin and Michael Maire and Serge Belongie and Lubomir Bourdev and Ross Girshick and James Hays and Pietro Perona and Deva Ramanan and C. Lawrence Zitnick and Piotr Dollar},
  year={2014}, eprint={1405.0312}, archivePrefix={arXiv}, url={https://arxiv.org/abs/1405.0312}
}
@article{li2019cococn,
  title={COCO-CN for Cross-Lingual Image Tagging, Captioning and Retrieval},
  author={Xirong Li and Chaoxi Xu and Xiaoxu Wang and Weiyu Lan and Zhengxiong Jia and Gang Yang and Jieping Xu},
  journal={IEEE Transactions on Multimedia}, volume={21}, number={9}, pages={2347--2360}, year={2019},
  eprint={1805.08661}, url={https://arxiv.org/abs/1805.08661}
}

@article{ravi2024sam2,
  title={SAM 2: Segment Anything in Images and Videos}, author={Ravi, Nikhila and others},
  year={2024}, eprint={2408.00714}, archivePrefix={arXiv}, url={https://arxiv.org/abs/2408.00714}}
@article{cho2025perceptionlm,
  title={PerceptionLM: Open-Access Data and Models for Detailed Visual Understanding},
  author={Cho, Jang Hyun and others}, year={2025}, eprint={2504.13180},
  archivePrefix={arXiv}, url={https://arxiv.org/abs/2504.13180}}

@article{cao2014cremad,
 title={CREMA-D: Crowd-sourced Emotional Multimodal Actors Dataset},
 author={Cao, Houwei and Cooper, David G. and Keutmann, Michael K. and Gur, Ruben C. and Nenkova, Ani and Verma, Ragini},
 journal={IEEE Transactions on Affective Computing},year={2014},volume={5},number={4},pages={377--390},
 doi={10.1109/TAFFC.2014.2336244},url={https://doi.org/10.1109/TAFFC.2014.2336244}}

@inproceedings{koizumi2023librittsr,
  title={LibriTTS-R: A Restored Multi-Speaker Text-to-Speech Corpus},
  author={Yuma Koizumi and Heiga Zen and Shigeki Karita and Yifan Ding and Kohei Yatabe and Nobuyuki Morioka and Michiel Bacchiani and Yu Zhang and Wei Han and Ankur Bapna},
  booktitle={Interspeech}, year={2023},
  eprint={2305.18802}, archivePrefix={arXiv}, url={https://arxiv.org/abs/2305.18802}
}
@misc{shi2020aishell3,
  title={AISHELL-3: A Multi-speaker Mandarin TTS Corpus and the Baselines},
  author={Yao Shi and Hui Bu and Xin Xu and Shaoji Zhang and Ming Li},
  year={2020}, eprint={2010.11567}, archivePrefix={arXiv},
  url={https://arxiv.org/abs/2010.11567}, note={arXiv revision 2, 2021; the OpenSLR page's 2015 BibTeX year is a typo}
}

@article{fonseca2022FSD50K,
 title={FSD50K: An Open Dataset of Human-Labeled Sound Events},
 author={Fonseca, Eduardo and Favory, Xavier and Pons, Jordi and Font, Frederic and Serra, Xavier},
 journal={IEEE/ACM Transactions on Audio, Speech, and Language Processing},
 volume={30}, pages={829--852}, year={2022}, doi={10.1109/TASLP.2021.3133208}
}

@inproceedings{thickstun2017learning,
 title={Learning Features of Music from Scratch},
 author={John Thickstun and Zaid Harchaoui and Sham M. Kakade},
 booktitle={International Conference on Learning Representations},
 year={2017},
 url={https://arxiv.org/abs/1611.09827}
}
@inproceedings{thickstun2018invariances,
 title={Invariances and Data Augmentation for Supervised Music Transcription},
 author={John Thickstun and Zaid Harchaoui and Dean P. Foster and Sham M. Kakade},
 booktitle={IEEE International Conference on Acoustics, Speech and Signal Processing},
 year={2018},
 url={https://arxiv.org/abs/1711.04845}
}

@misc{nasajsc2026crewobservations,author={{Earth Science and Remote Sensing Unit, NASA Johnson Space Center}},title={Crew Earth Observations Time-Lapse Videos},year={2026},url={https://eol.jsc.nasa.gov/BeyondThePhotography/CrewEarthObservationsVideos/},note={Specific 2011--2019 camera capture sequences and original NASA/JSC human descriptions; per-sequence source frame identities, UTC clocks, source snapshots and derivative hashes retained}}

@misc{wikimedia2023wikipedia,author={Wikipedia contributors},title={Wikipedia dump20231101,English and Chinese},year={2023},url={https://huggingface.co/datasets/wikimedia/wikipedia}}

Formal streaming metadata companion

releases/full-ft-v1/formal-index-v1/packets.jsonl.gz is a metadata-only index for the formal Colab reader. It contains every released packet and every task variant, with its original target/input roles and exact TAR member offsets and hashes. It includes the 40,000 human text continuation records. This prevents an initial full-media download just to discover tasks. INDEX.json binds it to media revision f472f980a6c95d3e9844f27e6045c8c09b21536a; the original media, rights ledgers, attribution and citations remain unchanged at that revision. The index contains 304,861 instructions and does not establish model-training readiness or an evaluation result.