nvidia/Nemotron-Image-Training-v3.5

Nemotron Super 3.5 VL Training Data

Dataset Description

This collection brings together training annotations selected from the data used for Nemotron Super 3.5 VL. It covers detection and grounding, scene-text understanding, counting, spatial and chart reasoning, visual puzzles, robotics, planning, and video question answering.

The collection contains 40 dataset subsets, comprising 6,540,914 annotation records and 29,991 bundled synthetic image files. The token estimates sum to approximately 13.99 billion.

The annotation subsets contain text labels, questions, answers, and, where applicable, generated reasoning, together with references to images or videos. For legal reasons, we unfortunately cannot release the images for most datasets. Users can download those images from the sources linked below.

This dataset is ready for commercial or non-commercial uses.

Dataset Owner

NVIDIA Corporation.

Dataset Creation Date

2026-10-12

Versioning:

Nemotron-Image-Training-v3.5

Previous Version(s): Nemotron-Image-Training-v3: Previous release in the Nemotron Image Training collection.

Relationship to Previous Version(s): Extension - This release supplements Nemotron-Image-Training-v3 and is intended to be used alongside it.

License/Terms of Use:

Use of the released annotations is governed by OpenMDW-1.1.

BabyVision image files are governed by OpenMDW-1.1, except that OpenMoji-derived images are governed by Creative Commons Attribution ShareAlike 4.0 International (CC BY-SA 4.0).

Where specified in the Dataset Composition table: Each user is responsible for checking the content of datasets and the applicable licenses and determining if suitable for the intended use.

Additional Information:

• Open Images V6 annotations (Google LLC), COCO annotations, and LVIS annotations: Creative Commons Attribution 4.0 International (CC BY 4.0).
• PACO, GroundCUA (ServiceNow), and MindCube: MIT License.
• PixMo-Points and PixMo-Count (Allen Institute for AI): Open Data Commons Attribution License (ODC-BY) v1.0.
• SAT: MIT License.
• RoboVQA (DeepMind Technologies Limited): Creative Commons Attribution 4.0 International (CC BY 4.0).

See the Dataset Composition table for the governing terms and additional information for each subset.

Intended Usage

The annotations are intended for training vision-language models.

Dataset Characterization

Collection and annotation methods: conversion of existing human or template annotations, model-generated annotations and reasoning, and NVIDIA procedural image generation. Filtering models select or reject examples; their inclusion in the Processing tags does not mean that their generated text is released.

Format: JSON Lines (.jsonl) annotations in the multimessage layout of Nemotron Image Training v3. Image assets retain their original formats.

Each line contains an id and messages. Message content contains text strings and media objects. Video references retain supported subclip ranges or frame positions.

{"id":"00000000-0000-4000-8000-000000000000","messages":[{"role":"user","content":[{"type":"image","image":"example.png"},"Describe the image."]},{"role":"assistant","content":["An example image."]}]}

The image field accepts a relative filename or an HTTP(S) URL. Image objects may include an optional image_sha256 field containing the 64-character lowercase SHA-256 of the original image bytes. PixMo image objects provide both the source URL and this hash:

{"type":"image","image":"https://example.org/image.jpg","image_sha256":"0000000000000000000000000000000000000000000000000000000000000000"}

The UUID is an export identifier. Record order, conversation text, media order, and supported timing fields are preserved. Empty redacted-thinking wrappers are removed as in the v3 exporter.

Dataset Composition

The table lists the 40 dataset subsets.

Dataset Name	Samples	Est. tokens	Data & Task Type	Source	Processing	Governing Terms
icdar_art_detection	55,320	22.44 M	image ocr grounding	ICDAR ArT	public qwen-3-vl-labels molmo-labels paddleocr-v5-labels sam-3-labels	Use of this dataset is governed by OpenMDW-1.1. Each user is responsible for checking the content of datasets and the applicable licenses and determining if suitable for the intended use.
bdd100k_detection	70,032	198.33 M	image grounding	BDD100K	public qwen-3-vl-labels molmo-labels paddleocr-v5-labels sam-3-labels	Use of this dataset is governed by OpenMDW-1.1. Each user is responsible for checking the content of datasets and the applicable licenses and determining if suitable for the intended use.
blip3_ocr_detection	294,856	153.30 M	image ocr grounding	BLIP3-OCR	public paddleocr-v5-labels	Use of this dataset is governed by OpenMDW-1.1. Each user is responsible for checking the content of datasets and the applicable licenses and determining if suitable for the intended use.
blip3_ocr_grounding	514,150	417.68 M	image ocr grounding	BLIP3-OCR	public paddleocr-v5-labels	Use of this dataset is governed by OpenMDW-1.1. Each user is responsible for checking the content of datasets and the applicable licenses and determining if suitable for the intended use.
openimages_detection	1,986,180	6.30 B	image grounding	Open Images	public rule-based-conversion	Use of this dataset is governed by OpenMDW-1.1. Each user is responsible for checking the content of datasets and the applicable licenses and determining if suitable for the intended use. Additional Information: Creative Commons Attribution 4.0 International (CC BY 4.0) for Open Images V6 annotations (Google LLC).
paco_ego4d_detection	15,667	87.76 M	image grounding	PACO / Ego4D	public rule-based-conversion	Use of this dataset is governed by OpenMDW-1.1. Each user is responsible for checking the content of datasets and the applicable licenses and determining if suitable for the intended use. Additional Information: MIT License for PACO; Creative Commons Attribution 4.0 International (CC BY 4.0) for LVIS annotations.
paco_lvis_detection	45,802	198.33 M	image grounding	PACO / LVIS	public rule-based-conversion	Use of this dataset is governed by OpenMDW-1.1. Each user is responsible for checking the content of datasets and the applicable licenses and determining if suitable for the intended use. Additional Information: MIT License for PACO; Creative Commons Attribution 4.0 International (CC BY 4.0) for LVIS annotations.
pixmo_points_pointing	561,004	626.27 M	image grounding	PixMo-Points	public rule-based-conversion	Use of this dataset is governed by OpenMDW-1.1. Each user is responsible for checking the content of datasets and the applicable licenses and determining if suitable for the intended use. Additional Information: Open Data Commons Attribution License (ODC-BY) v1.0 for PixMo-Points (Allen Institute for AI).
pixmo_count_pointing	31,993	45.61 M	image grounding	PixMo-Count	public rule-based-conversion	Use of this dataset is governed by OpenMDW-1.1. Each user is responsible for checking the content of datasets and the applicable licenses and determining if suitable for the intended use. Additional Information: Open Data Commons Attribution License (ODC-BY) v1.0 for PixMo-Count (Allen Institute for AI).
sat_spatial_qa	172,384	54.94 M	multi-image qa	SAT	public rule-based-conversion	Use of this dataset is governed by OpenMDW-1.1. Each user is responsible for checking the content of datasets and the applicable licenses and determining if suitable for the intended use. Additional Information: MIT License for SAT.
coco_count_spatial_qa	14,777	5.21 M	image qa	COCO	public rule-based-conversion	Use of this dataset is governed by OpenMDW-1.1. Each user is responsible for checking the content of datasets and the applicable licenses and determining if suitable for the intended use. Additional Information: Creative Commons Attribution 4.0 International (CC BY 4.0) for COCO annotations.
groundcua_grounding	999,528	1.98 B	image grounding	GroundCUA	public qwen-3-vl-labels	Use of this dataset is governed by OpenMDW-1.1. Each user is responsible for checking the content of datasets and the applicable licenses and determining if suitable for the intended use. Additional Information: MIT License for GroundCUA (ServiceNow).
mindcube_spatial_qa	10,000	22.11 M	multi-image qa	MindCube	public rule-based-conversion	Use of this dataset is governed by OpenMDW-1.1. Each user is responsible for checking the content of datasets and the applicable licenses and determining if suitable for the intended use. Additional Information: MIT License for MindCube.
robovqa	793,538	1.10 B	video qa	RoboVQA	public rule-based-conversion	Use of this dataset is governed by OpenMDW-1.1. Each user is responsible for checking the content of datasets and the applicable licenses and determining if suitable for the intended use. Additional Information: Creative Commons Attribution 4.0 International (CC BY 4.0) for RoboVQA (DeepMind Technologies Limited).
pixmo_count_qa	34,295	49.61 M	image qa	PixMo-Count	public rule-based-conversion	Use of this dataset is governed by OpenMDW-1.1. Each user is responsible for checking the content of datasets and the applicable licenses and determining if suitable for the intended use. Additional Information: Open Data Commons Attribution License (ODC-BY) v1.0 for PixMo-Count (Allen Institute for AI).
hopchain_fintabnet	56,479	211.24 M	image qa reasoning	FinTabNet	public qwen-3.5-labels sam-3-labels	Use of this dataset is governed by OpenMDW-1.1. Each user is responsible for checking the content of datasets and the applicable licenses and determining if suitable for the intended use.
hopchain_qa	7,328	50.67 M	image qa reasoning	Common Crawl (human-annotated)	public qwen-3.5-labels sam-3-labels	Use of this dataset is governed by OpenMDW-1.1. Each user is responsible for checking the content of datasets and the applicable licenses and determining if suitable for the intended use.
deltaprompt_chart_reasoning	10,440	66.53 M	image qa reasoning	ArxivQA	public qwen-3-vl-labels qwen-3.5-labels	Use of this dataset is governed by OpenMDW-1.1. Each user is responsible for checking the content of datasets and the applicable licenses and determining if suitable for the intended use.
arxiv_chart_qa_01	212,056	551.01 M	image qa	arXiv chart sources	public qwen-3.5-labels	Use of this dataset is governed by OpenMDW-1.1. Each user is responsible for checking the content of datasets and the applicable licenses and determining if suitable for the intended use.
chartnet_qa_01	24,750	52.40 M	image qa reasoning	ChartNet	public qwen-3.5-labels	Use of this dataset is governed by OpenMDW-1.1.
chartnet_qa_01_difficulty	9,498	37.50 M	image qa reasoning	ChartNet	public qwen-3.5-labels	Use of this dataset is governed by OpenMDW-1.1.
chartnet_qa_02	66,467	144.21 M	image qa reasoning	ChartNet	public qwen-3.5-labels	Use of this dataset is governed by OpenMDW-1.1.
chartnet_qa_02_difficulty	25,516	101.05 M	image qa reasoning	ChartNet	public qwen-3.5-labels	Use of this dataset is governed by OpenMDW-1.1.
chartnet_qa_03	42,704	90.38 M	image qa reasoning	ChartNet	public qwen-3.5-labels	Use of this dataset is governed by OpenMDW-1.1.
chartnet_qa_03_difficulty	16,929	67.36 M	image qa reasoning	ChartNet	public qwen-3.5-labels	Use of this dataset is governed by OpenMDW-1.1.
chartnet_qa_04	58,280	123.46 M	image qa reasoning	ChartNet	public qwen-3.5-labels	Use of this dataset is governed by OpenMDW-1.1.
chartnet_qa_04_difficulty	22,747	90.34 M	image qa reasoning	ChartNet	public qwen-3.5-labels	Use of this dataset is governed by OpenMDW-1.1.
chartnet_qa_05	26,547	57.84 M	image qa reasoning	ChartNet	public qwen-3.5-labels	Use of this dataset is governed by OpenMDW-1.1.
chartnet_qa_05_difficulty	10,454	40.20 M	image qa reasoning	ChartNet	public qwen-3.5-labels	Use of this dataset is governed by OpenMDW-1.1.
chartnet_qa_06_difficulty	29,909	118.85 M	image qa reasoning	ChartNet	public qwen-3.5-labels	Use of this dataset is governed by OpenMDW-1.1.
chartverse_reasoning_01	19,289	77.02 M	image qa reasoning	ChartVerse	public qwen-3.5-labels	Use of this dataset is governed by OpenMDW-1.1.
chartverse_reasoning_02	16,233	64.45 M	image qa reasoning	ChartVerse	public qwen-3.5-labels	Use of this dataset is governed by OpenMDW-1.1.
arxiv_chart_qa_02	113,428	316.09 M	image qa	arXiv chart sources	public qwen-3.5-labels	Use of this dataset is governed by OpenMDW-1.1.
chartverse_perception_qa	15,194	50.73 M	image qa reasoning	ChartVerse	public qwen-3.5-labels	Use of this dataset is governed by OpenMDW-1.1.
babyvision_reasoning	135,146	331.27 M	image qa reasoning	NVIDIA procedural puzzle images	nvidia-generated qwen-3.5-labels	Use of this dataset is governed by OpenMDW-1.1, except that use of the OpenMoji derived images is governed by Creative Commons Attribution ShareAlike 4.0 International.
vision_language_reasoning	9,808	30.07 M	image qa reasoning	VISION_LANGUAGE	public qwen-3.5-labels	Use of this dataset is governed by OpenMDW-1.1.
visualpuzzles_reasoning	635	2.08 M	image qa reasoning	VisualPuzzles	public qwen-3.5-labels	Use of this dataset is governed by OpenMDW-1.1.
deepvision_reasoning	9,818	33.75 M	image qa reasoning	DeepVision	public qwen-3.5-labels	Use of this dataset is governed by OpenMDW-1.1. Each user is responsible for checking the content of datasets and the applicable licenses and determining if suitable for the intended use.
perception_test_qa_01	1,388	5.18 M	video qa	Perception Test	public qwen-3.5-labels	Use of this dataset is governed by OpenMDW-1.1. Each user is responsible for checking the content of datasets and the applicable licenses and determining if suitable for the intended use.
perception_test_qa_02	345	1.56 M	video qa	Perception Test	public qwen-3.5-labels	Use of this dataset is governed by OpenMDW-1.1. Each user is responsible for checking the content of datasets and the applicable licenses and determining if suitable for the intended use.
Total annotation records (40 subsets)	6,540,914	13.99 B

Samples counts annotation records. Image-file counts are listed separately in Dataset Quantification and are not added to the annotation total. Counts across subsets have not been deduplicated.

Est. tokens uses training-statistics values. Of the 40 token counts, 34 are estimated and six are marked exact under their accounting method.

Tokenizer: Nemotron 3.5 BPE (similar to the Nemotron 3 Super tokenizer), with a 131,072-token base vocabulary. Counting uses the Nemotron 3.5 nemotron6-moe chat format and multimodal special-token mappings, which differ from the public reference.

Token accounting: Padding-free lengths of training-preprocessed multimodal conversations, including prompt and response text, reasoning spans, chat-format control tokens, and image or sampled-video-frame embedding tokens. Media token lengths are calculated from image dimensions and video frame-sampling settings using the training encoder's expansion rules. Text is tokenized after applying the chat template, with historical thinking retained and no added generation prompt. Dataset totals are exact sums for completed censuses or estimates extrapolated from observed sequence lengths to the source dataset population, including stratified estimates where used. These are source training-statistics counts, rather than a recount of the final released JSONLs.

Tag Legend

• image: A referenced image, or image files included with BabyVision.
• multi-image: Samples can reference multiple images; includes frame sequences.
• video: Samples reference videos.
• ocr: Text-in-image supervision.
• grounding: Coordinates, bounding boxes, regions or pointing supervision; also covers detection.
• qa: Question-answer supervision. Chart, visual, spatial, counting and planning QA share this tag.
• multi-qa: Multiple question-answer turns within one sample.
• reasoning: Reasoning traces.
• public: Derived from a public upstream corpus or benchmark; its access and usage terms still apply.
• nvidia-generated: NVIDIA-created source images, including the procedural BabyVision images and their paired annotations.
• rule-based-conversion: Deterministic formatting, coordinate conversion or template expansion of existing supervision. Any model-generation stage is indicated separately where verified.
• <model>-<version>-labels: Model used for labeling.

Model tags group checkpoints by family. The Source column holds upstream links. Documented rule conversions have a separate rule-based-conversion tag. The absence of a teacher-model tag does not establish that a dataset's entire upstream history is model-free.

Dataset Quantification

• Dataset subsets: 40. The BabyVision annotations and image sets are grouped under babyvision_reasoning.
• Aggregate token estimate: 13.99 B.
• BabyVision includes 29,991 bundled synthetic image files; 19,321 are referenced by the released annotations.
• Download size (JSONLs only): 23.71 GB (22.08 GiB).
• Download size (bundled image tar archives): 4.38 GB (4.08 GiB).
• Download size (JSONLs and included images): 28.09 GB (26.16 GiB).
• Referenced-media volume: approximately 2.11 TB (1.92 TiB).
• Referenced image volume: approximately 1.59 TB.
• Referenced video volume: approximately 522 GB.
• Release record count: 6,540,914 across the 40 released JSONL files, counted without deduplication.

Media volume sums encoded image and video bytes referenced by the 40 annotation subsets. Repeated references within and across subsets are counted each time; video bytes refer to full encoded video files. These figures use the training-statistics snapshot dated 2026-09-19. Image totals combine exact measurements and estimates; RoboVQA's video volume is estimated.

Loading the Data

For each subdataset, repo contains a JSONL file at <dataset>/<dataset>.jsonl and a component dataset card at <dataset>/README.md. See the component cards for the applicable upstream-media download and extraction instructions.

Local media paths in the JSONL resolve relative to the release repository root. HTTP(S) references point directly to the original downloadable images. Verified upstream filesystem paths are already written into the annotations. For datasets with images embedded in upstream files, follow the component card's extraction instructions to write images directly to the referenced filenames; no separate filename mapping table is supplied.

Synthetic BabyVision image tar shards are colocated under babyvision_reasoning/media/. Extract them into babyvision_reasoning/ from the release repository root, as described in the BabyVision component card:

for archive in babyvision_reasoning/media/*.tar; do
  tar -xf "$archive" -C babyvision_reasoning
done

Dataset Limitations

Generated annotations can contain factual or reasoning errors.

Ethical Considerations:

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. Developers should work with their internal developer teams to ensure this dataset meets requirements for the relevant industry and use case and addresses unforeseen product misuse.

Please report quality, risk, security vulnerabilities or NVIDIA AI Concerns here.