GPAI Ledger › GPAI Training Transparency tracker (AI Accountability Lab (AIAL)) › Capture 12 Sep 2026

NVIDIA_Nemotron_3_5_2026_09_08 — capture 20260912T062147Z

Filed under AI Accountability Lab (AIAL) — GPAI Training Transparency tracker, the source this project captured it from; the document itself is the filing of the model named above.

Verify: sha256sum 69ae78125314d050a048570dbcedf2f9efc74bbae73cb508725f20fb686b6c56.pdf must equal the hash above (the filename IS the expected hash); ots verify 69ae78125314d050a048570dbcedf2f9efc74bbae73cb508725f20fb686b6c56.pdf.20260912T062147Z.ots -f 69ae78125314d050a048570dbcedf2f9efc74bbae73cb508725f20fb686b6c56.pdf (opentimestamps.org) proves the bytes existed no later than the attestation time — an upper bound on the capture time; the fetch time above is the archive's own record (a freshly captured proof reports 'pending' here: the calendars anchor within hours, but this archive only upgrades the stored proof to its anchor on a later run, so expect a day or two). ots verify needs a local Bitcoin Core node (a pruned one is fine); without one, ots info on the proof prints the attesting block height and merkle path to check on any block explorer.

Extracted text

Machine-extracted text (layout may be lost; the authoritative content is the stored file above).

Public  Summary  of  Training  Content  for
NVIDIA

Nemotron

Nano

v2

12B

VL

 1.  General  information
1.1.  Provider  identification  Provider  name  and  contact  details
NVIDIA  Corporation  2788  San  Tomas  Expressway   Santa  Clara,  CA  95051,  USA
Authorised  representative  name  and  contact  details
NVIDIA  GmbH  Adenauerstraße  20  A4  52146  Würselen,  Germany  AI_Governance@nvidia.com

1.2.  Model  identification  Versioned  model  name(s)  NVIDIA-Nemotron-Nano-12B-v2-VL-BF16  (https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16)  NVIDIA-Nemotron-Nano-12B-v2-VL-FP8  (https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-FP8)  NVIDIA-Nemotron-Nano-12B-v2-VL-NVFP4-QAD  (https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-NVFP4-QAD)
Model  dependencies  Developed  from  NVIDIA-Nemotron-Nano-12B-v2.
Date  of  placement  of  the  model  on  the  Union  market
27  October  2025

1.3.  Modalities,  overall  training  data  size  and  other  characteristics  Modality  Training  data  size  Types  of  content
[X]  Text  [  ]  Less  than  1  billion  tokens  [X]  1  billion  to  10  trillion  tokens  []  More  than  10  trillion  tokens
Instruction  tuning,  reasoning  and  question  answering,  source  code,  Optical  Character  Recognition  (OCR)-derived  text,  captions,  document  text,  and  synthetic  text.
[X]  Image  [  ]  Less  than  1  million  images  [X]  1  million  to  1  billion  images  [  ]  More  than  1  billion  images
Natural  and  digital  images,  business  documents,  diagrams,  charts,  tables,  scientific  documents,  text-bearing  images,  and  synthetically  rendered  OCR  material.
[  ]  Audio  ☐  Less  than  10  000  hours  ☐  10,000  to1  million  hours  ☐  More  than  1  million  hours

[X]  Video  [  ]  Less  than  10.000  hours  [X]  10,000  to  1  billion  hours  []  More  than  1  billion  hours
Audiovisual  content,  including  instructional  and  activity  videos  and  general-interest  audiovisual  material,  captions,  question  and  answers.
[  ]  Other   N/A    Latest  date  of  data  acquisition/collection  for  model  training
2025
1
Description  of  linguistic  characteristics
The  training  corpus  includes  a  variety  of  languages,  including  English,  Chinese,  German,  Spanish,  French,  Italian,  Japanese,  Korean,  Dutch,  and  Portuguese.
Other  relevant  characteristics

Additional  comments  (optional)

2.  List  of  data  sources
2.1.  Publicly  available  datasets  Have  you  used  publicly  available  datasets  to  train  the  model?
[X]  Yes     [  ]  No
If  yes,  specify  the  modality(ies)  of  the  content  covered  by  the  datasets  concerned:
[X]  Text     [X]  Image     [X]  Video     [  ]  Audio     [  ]  Other
List  of  large  publicly  available  datasets:
See  Appendix  A.1       General  description  of  other  publicly  available  datasets  not  listed  above:
Text  datasets  including  instruction  tuning,  reasoning,  question  answering,  safety  alignment,  and  function  calling.  Image-and-text  datasets  containing  OCR,  document  and  image  reasoning,  visual  question  answering  (VQA),  captioning,  visual  instruction  tuning,  safety,  grounding,  and  classification.  Video-and-text  datasets  including  video  question  answering,  captioning,  temporal  and  spatial  reasoning,  action  recognition,  and  safety.
Additional  comments  (optional)

2.2.  Private  non-publicly  available  datasets  obtained  from  third  parties  2.2.1.  Datasets  commercially  licensed  by  rightsholders  or  their  representatives  Have  you  concluded  transactional  commercial  licensing  agreement(s)  with  rightsholder(s)  or  with  their  representatives?
[]  Yes     [X]  No
If  yes,  specify  the  modality(ies)  of  the  content  covered  by  the  datasets  concerned:
[]  Text     [  ]  Image     [  ]  Video     [  ]  Audio     [  ]  Other

 2.2.2.  Private  datasets  obtained  from  other  third  parties  Have  you  obtained  private  datasets  from  third  parties  that  are  not  licensed  as  described  in  Section  2.2.1,  such  as  data  obtained  from  providers  of  private  databases,  or  data  intermediaries?
[X]  Yes     [  ]  No
2
If  yes,  specify  the  modality(ies)  of  the  content  covered  by  the  datasets  concerned:
[X]  Text     [X]  Image     [  ]  Video     [  ]  Audio     [  ]  Other
If  publicly  known,  list  private  datasets  obtained  from  other  third  parties:
N/A
General  description  of  non-publicly  known  private  datasets  obtained  from  third  parties
Invoice  documents  with  question-answer  annotations.
Additional  comments  (optional)

2.3.  Data  crawled  and  scraped  from  online  sources  Were  crawlers  used  by  the  provider  or  on  behalf  of?
[X]  Yes     [  ]  No
Crawler  Names/Identifier:  NVIDIA  Crawler  Identifier:  nvidiabot  Additional  third-party  crawler  identifiers  unknown.
Purposes  of  the  crawler(s):   Acquire  or  curate  images  for  captioning  and  enterprise  visual  question  answering  (VQA);  derive  dense  image-caption  data  from  Common  Crawl  materials;  and  create  video  captioning,  question-answering,  and  reasoning  datasets.
General  description  of  crawler  behaviour
Automated  crawlers  were  used  to  collect  publicly  available  media  from  selected  websites.  The  crawlers  navigated  publicly  accessible  pages,  extracted  relevant  links  and  associated  metadata,  and  collected  eligible  content  for  model  training  and  evaluation.
Period  of  data  collection  2023-2025
Comprehensive  description  of  the  type  of  content  and  online  sources  crawled:
Natural  and  digital  images  used  for  captioning  and  VQA,  plus  videos  used  for  captioning,  question  answering,  and  reasoning.
Type  of  modality  covered:   [X]  Text     [X]  Image     [X]  Video     [  ]  Audio     [  ]  Other
Summary  of  the  most  relevant  domain  names  crawled:

Additional  comments  (optional)

2.4.  User  data  Was  data  from  user  interactions  with  the  AI  model  (e.g.  user  input  and  prompts)  used  to  train  the  model?
[]  Yes     [  X]  No.
3
Was  data  collected  from  user  interactions  with  the  provider’s  other  services  or  products  used  to  train  the  model?
[]  Yes     [  X]  No.
If  yes,  provide  a  general  description  of  the  provider’s  services  or  products  that  were  used  to  collect  the  user  data:

Type  of  modality  covered:
Additional  comments  (optional)

2.5.  Synthetic  data  Was  synthetic  AI-generated  data  created  by  the  provider  or  on  their  behalf  to  train  the  model?
[X]  Yes     [  ]  No
If  yes,  modality  of  the  synthetic  data:
[X]  Text     [X]  Image     [  ]  Video     [  ]  Audio     [  ]  Other
If  yes,  specify  the  general-purpose  AI  model(s)  used  to  generate  the  synthetic  data  if  available  on  the  market:
DeepSeek-R1-0528;  Qwen3-235B-A22B;  Qwen3-30B;  Qwen2.5-32B-Instruct-AWQ;  Qwen2.5-14B-Instruct;  Qwen2.5-VL-32B-Instruct
Information  about  other  AI  models,  including  provider’s  own  AI  model(s)  not  available  on  the  market,  used  to  generate  synthetic  data  to  train  the  model  to  which  this  Summary  applies:

Additional  comments  (optional)

4
2.6.  Other  sources  of  data  Have  data  sources  other  than  those  described  in  Sections  2.1  to  2.5  been  used  to  train  the  model?

[]  Yes     [X]  No
If  yes,  provide  a  narrative  description  of  these  data  sources  and  the  data:

Additional  comments  (optional)

3.  Data  processing  aspects
3.1.  Respect  of  reservation  of  rights  from  text  and  data  mining  exception  or  limitation  Are  you  a  Signatory  to  the  Code  of  Practice  for  general-purpose  AI  models  that  includes  commitments  to  respect  reservations  of  rights  from  the  TDM  exception  or  limitation?
[]  Yes     [X]  No
Describe  the  measures  implemented  before  model  training  to  respect  reservations  of  rights  from  the  TDM  exception  or  limitation  before  and  during  data  collection,  including  the  opt-out  protocols  and  solutions  honoured  by  the  provider  or,  as  applicable,  by  third  parties  from  which  datasets  have  been  obtained:
NVIDIA  implements  measures  to  respect  EU  rights  reservations  relevant  to  text  and  data  mining  by:  (1)  respecting  robots.txt  directives  at  the  domains  accessed  where  those  directives  constituted  a  reservation  of  rights  under  Article  4(3)  of  Directive  (EU)  2019/790;  and  (2)  filtering  datasets  on  any  actionable  metadata  identifiers  provided  by  rightsholders.
Additional  comments  (optional)

3.2.  Removal  of  illegal  content  General  description  of  measures  taken
Image-based  datasets  were  scanned  against  known  child  sexual  abuse  material  (CSAM),  with  the  inventory  recording  passed  compliance  scans  or  reuse  of  prior  scans  for  identical  source  images  where  applicable.  Safety-alignment  datasets,  rule-based  filtering,  and  model-based  filtering  were  also  used.   Datasets  were  subject  to  content-safety  checks  appropriate  to  their  source  and  modality.  Web-derived  image-text  data  underwent  automated  NSFW  screening  designed  to  identify  explicit  imagery  and  profane  or  inappropriate  accompanying  text,  with  flagged  samples  removed  before  using  the  data  for  training.

5
3.3.  Other  information  (optional)  Other  relevant  information  about  data  processing  (optional)
N/A

6

7
Appendix  A.  Datasets

A.1.  Publicly  available  datasets  A.1.1.  Text  datasets
Qualifying

source

lineage

Public

source

STEM

MMLU-aux

(subset)
 Source

MetaMathQA

(CoT)
 Source

8

A.1.2.  Image-and-text  datasets

Qualifying

source

lineage

Public

source

Open

Images

(captioning

and

bounding-box-derived

subsets)
 Source

Wikipedia-based

Image

Text

(WIT)
 Source

arXiv-derived

OCR

and

equation

data
 Source

arXiv-derived

OCR

and

equation

data
 Source

arXiv-derived

OCR

and

equation

data
 Source

arXiv-derived

OCR

and

equation

data
 Source

arXiv-derived

OCR

and

equation

data
 Source

arXiv-derived

OCR

and

equation

data
 Source

9

A.1.2.  Image-and-text  datasets

Qualifying

source

lineage

Public

source

arXiv-derived

OCR

and

equation

data
 Source

arXiv-derived

OCR

and

equation

data
 Source

arXiv-derived

OCR

and

equation

data
 Source

arXiv-derived

OCR

and

equation

data
 Source

Wikimedia

enterprise

HTML-derived

OCR

data
 Source

Wikimedia

enterprise

HTML-derived

OCR

data
 Source

Wikimedia

enterprise

HTML-derived

OCR

data
 Source

Wikimedia

enterprise

HTML-derived

OCR

data
 Source

10

A.1.2.  Image-and-text  datasets

Qualifying

source

lineage

Public

source

Wikimedia

enterprise

HTML-derived

OCR

data
 Source

Wikimedia

enterprise

HTML-derived

OCR

data
 Source

Wikimedia

enterprise

HTML-derived

OCR

data
 Source

Wikimedia

enterprise

HTML-derived

OCR

data
 Source

Wikimedia

enterprise

HTML-derived

OCR

data
 Source

Wikimedia

enterprise

HTML-derived

OCR

data
 Source

Wikimedia

enterprise

HTML-derived

OCR

data
 Source

Wikimedia

enterprise

HTML-derived

OCR

data
 Source

11

A.1.2.  Image-and-text  datasets

Qualifying

source

lineage

Public

source

Wikimedia

enterprise

HTML-derived

OCR

data
 Source

Wikimedia

enterprise

HTML-derived

OCR

data
 Source

Wikimedia

enterprise

HTML-derived

OCR

data
 Source

Wikimedia

enterprise

HTML-derived

OCR

data
 Source

Wikimedia

enterprise

HTML-derived

OCR

data
 Source

Wikimedia

enterprise

HTML-derived

OCR

data
 Source

Wikimedia

enterprise

HTML-derived

OCR

data
 Source

Wikimedia

enterprise

HTML-derived

OCR

data
 Source

12

A.1.2.  Image-and-text  datasets

Qualifying

source

lineage

Public

source

Open

Images

(captioning

and

bounding-box-derived

subsets)
 Source

Open

Images

(captioning

and

bounding-box-derived

subsets)
 Source

Open

Images

(captioning

and

bounding-box-derived

subsets)
 Source

arXiv-derived

OCR

and

equation

data
 Source

13

A.1.3.  Video-and-text  datasets

Qualifying

source

lineage

Public

source

ActivityNet
 Source

ActivityNet
 Source

ActivityNet
 Source

ActivityNet
 Source

Kinetics
 Source

EgoExoLearn
 Source

Kinetics
 Source

ActivityNet
 Source

14

A.1.3.  Video-and-text  datasets

Qualifying

source

lineage

Public

source

HACS
 Source

ActivityNet
 Source

EgoExoLearn
 Source

EgoExoLearn
 Source

HC-STVG
 Source

HC-STVG
 Source

15
Public  Summary  of  Training  Content  for  NVIDIA
Nemotron

3

and

3.5

Family

Version  1.0   Last  update:  10/08/2026
1.  General  information
1.1.  Provider  identification

Provider  name  and  contact  details  NVIDIA  Corporation  2788  San  Tomas  Expressway,  Santa  Clara,  CA  95051,  USA
Authorised  representative  name  and  contact  details
NVIDIA  GmbH  Adenauerstraße  20  A4  52146  Würselen,  Germany  AI_Governance@nvidia.com
1.2.  Model  identification

Versioned  model  name(s)  Nemotron  3  Nano  30B  A3B  (v1.0)  •  NVIDIA-Nemotron-3-Nano-30B-A3B-Base-BF16  —  https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-Base-BF16 •  NVIDIA-Nemotron-3-Nano-30B-A3B-BF16  —  https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 •  NVIDIA-Nemotron-3-Nano-30B-A3B-FP8  —  https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8

•  NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4  —  https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4  Nemotron  3  Nano  Omni  30B  A3B  Reasoning  (v1.0)  •  Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16  —  https://huggingface.co/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16 •  Nemotron-3-Nano-Omni-30B-A3B-Reasoning-FP8  —  https://huggingface.co/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-FP8 •  Nemotron-3-Nano-Omni-30B-A3B-Reasoning-NVFP4  —  https://huggingface.co/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-NVFP4  Nemotron  3  Super  120B  A12B  (v1.0  GA  /  v1.0  Base)  •  NVIDIA-Nemotron-3-Super-120B-A12B-Base-BF16  —  https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-Base-BF16 •  NVIDIA-Nemotron-3-Super-120B-A12B-BF16  —  https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16 •  NVIDIA-Nemotron-3-Super-120B-A12B-FP8  —

https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-FP8 •  NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4  —  https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4  Nemotron  3  Ultra  550B  A55B  (v1.0  GA)  •  NVIDIA-Nemotron-3-Ultra-550B-A55B-Base-BF16  —  https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-Base-BF16 •  NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16  —  https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 •  NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4  —  https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4 •  NVIDIA-Nemotron-3-Ultra-550B-A55B-GenRM  —  https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-GenRM   Nemotron  3.5  •  https://huggingface.co/nvidia/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 •  https://huggingface.co/nvidia/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-Base-BF16 •

https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 •  https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4-DSpark •  https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4-DFlash
Model  dependencies  Nemotron  3.5  Lightning  30B  A3B  is  developed  from  NVIDIA-Nemotron-3-Nano-30B-A3B  (Mamba2-Transformer  hybrid  MoE  backbone).   Nano  30B  A3B,  Super  120B  A12B  and  Ultra  550B  A55B  are  not  fine-tunes  or  modifications  of  a  previously  placed  general-purpose  AI  model.  FP8  and  NVFP4  variants  are  post-training  quantisations  of  the  corresponding  BF16  checkpoint  and  were  not  separately  trained.   Nano  Omni  30B  A3B  Reasoning  is  developed  from  NVIDIA-Nemotron-3-Nano-30B-A3B  (Mamba2-Transformer  hybrid  MoE  backbone).
Date  of  placement  of  the  model  on  the  Union  market
15  December  2025:  Nemotron  3  Nano  30B  A3B  (all  variants)   11  March  2026:  Nemotron  3  Super  120B  A12B  (aligned,  FP8,  NVFP4).    28  April  2026:  Nemotron  3  Nano  Omni  30B  A3B  Reasoning  (all  variants)

4  June  2026:  Nemotron  3  Ultra  550B  A55B  (all  variants)   11  August  2026:  Nemotron  3.5  Lightning  30B  A3B  (all  variants)
1.3.  Modalities,  overall  training  data  size  and  other
characteristics

Modality  Training  data  size  Types  of  content
[X]  Text  [  ]  Less  than  1  billion  tokens  [  ]  1  billion  to  10  trillion  tokens  [X]  More  than  10  trillion  tokens
Web  pages,  dialogue,  articles  and  other  written  materials;  source  code  in  43  programming  languages;  mathematical  and  scientific  text;  legal  and  regulatory  text;  biomedical  text;  competitive  programming  and  software  engineering;  financial  filings;  academic  papers  and  patents;  question  answering,  instruction  tuning,  tool  use,  function  and  tool  calling,  structured-output  and  alignment  data;  reasoning  traces;  and  synthetically  generated  and  synthetically  rephrased  text.
[X]  Image  [  ]  Less  than  1  million  images  [X]  1  million  to  1  billion  images   [  ]  More  than  1  billion  images
Applies  to  Nemotron  3  Nano  Omni  Only:  Natural  and  digital  images,  business  and  scientific  documents,  PDFs,  charts,  plots,  diagrams,  tables,  screenshots  and  graphical  user  interfaces,  text-bearing  images  and
Modality  Training  data  size  Types  of  content
OCR  material,  medical  images,  geometry  and  mathematics  figures,  visual  question  answering,  captions,  and  synthetically  rendered  or  re-captioned  imagery.
[X]  Audio  [  ]  Less  than  10,000  hours  [X]  10,000  to  1  million  hours  [  ]  More  than  1  million  hours
Applies  to  Nemotron  3  Nano  Omni  Only:   Speech  and  general  audio  for  automatic  speech  recognition,  speech-to-text  chat,  audio  question  answering  and  audio  captioning,  including  text-to-speech-synthesised  material.
[X]  Video  [  ]  Less  than  10,000  hours  [X]  10,000  to  1  million  hours  [  ]  More  than  1  million  hours
Applies  to  Nemotron  3  Nano  Omni  Only:   Audiovisual  content  including  instructional  and  activity  video,  general-interest  video,  egocentric  and  procedural  footage,  with  captions,  question-answer  pairs  and  temporal/spatial  reasoning  annotations.
[  ]  Other   Not  Applicable
Latest  date  of  data  acquisition/collection  for  model  training
August  2026
Description  of  linguistic  characteristics  The  pre-training  corpus  includes  English,  Arabic,  Chinese,  Czech,  Danish,  Dutch,  Finnish,  French,  German,  Hebrew,  Hindi,

Italian,  Japanese,  Korean,  Polish,  Portuguese,  Russian,  Spanish,  Swedish  and  Thai.
Other  relevant  characteristics  The  corpus  spans  domains  including  legal  and  regulatory,  mathematics,  science,  medicine,  finance,  software  engineering  and  general  knowledge.  A  substantial  proportion  is  synthetically  generated  or  synthetically  rephrased.
Additional  comments  (optional)
2.  List  of  data  sources
2.1.  Publicly  available  datasets

Have  you  used  publicly  available  datasets  to  train  the  model?
[X]  Yes     [  ]  No
If  yes,  specify  the  modality(ies)  of  the  content  covered  by  the  datasets  concerned:
[X]  Text     [X]  Image     [X]  Video     [X]  Audio     [  ]  Other
List  of  large  publicly  available  datasets:  Common  Crawl  (https://commoncrawl.org/);  CC-NEWS  (https://commoncrawl.org/blog/news-dataset-available);  Wikimedia  (https://dumps.wikimedia.org/);  FineWeb-2  (https://huggingface.co/datasets/HuggingFaceFW/fineweb-2);  finepdfs  (https://huggingface.co/datasets/HuggingFaceFW/finepdfs);  Essential-Web  v1.0  (https://huggingface.co/datasets/EssentialA

I/essential-web-v1.0);  The  Common  Pile  v0.1  (https://huggingface.co/common-pile);  MegaMath  (https://huggingface.co/datasets/LLM360/MegaMath);  FineMath  (https://huggingface.co/datasets/HuggingFaceTB/finemath);  arXiv  (https://info.arxiv.org/help/bulk_data/index.html);  peS2o  (https://huggingface.co/datasets/allenai/peS2o);  Stack  Exchange  Data  Dump  (https://archive.org/details/stackexchange);  Reddit  Dataset  (https://files.pushshift.io/reddit/);  SEC  EDGAR  (https://www.sec.gov/search-filings);  Court  Listener  (https://www.courtlistener.com/help/api/bulk-data/);  LMSYS-Chat-1M  (https://huggingface.co/datasets/lmsys/lmsys-chat-1m);  WildChat-1M  (https://huggingface.co/datasets/allenai/WildChat-1M);  MiraData;  laion-disco-12M;  OpenImages  (https://storage.googleapis.com/openimages/web/index.html);  Conceptual  Captions  3M;  MINT-1T.    General  description  of  other  publicly  available  datasets  not  listed  above:
Text:  instruction  tuning,  reasoning,  question  answering,  mathematics,  competitive  programming  and  software  engineering,  function  and  tool  calling,  safety  alignment,  translation  and  multilingual  corpora,  and  academic/biomedical/legal/financial  document  collections.   Image-and-text:  OCR  and  document  understanding,  chart/table/diagram
reasoning,  visual  question  answering,  captioning,  content  for  visual  grounding,  computer-use  and  GUI  interaction,  visual  instruction  tuning,  geometry  and  vision-critical  STEM,  and  classification.   Video-and-text:  video  question  answering,  captioning,  temporal  and  spatial  reasoning,  procedural  and  egocentric  activity  understanding.   Audio-and-text:  automatic  speech  recognition,  speech  instruction  following,  audio  captioning  and  audio  question  answering.
Additional  comments  (optional)
2.2.  Private  non-publicly  available  datasets  obtained  from  third
parties

2.2.1.  Datasets  commercially  licensed  by  rightsholders  or  their
representatives

Have  you  concluded  transactional  commercial  licensing  agreement(s)  with  rightsholder(s)  or  with  their  representatives?
[X]  Yes     [  ]  No
If  yes,  specify  the  modality(ies)  of  the  content  covered  by  the  datasets  concerned:
[X]  Text     [X]  Image     [X]  Video     [X]  Audio     [  ]  Other

2.2.2.  Private  datasets  obtained  from  other  third  parties

Have  you  obtained  private  datasets  from  third  parties  that  are  not  licensed  as  described  in  Section  2.2.1,  such  as  data  obtained  from  providers  of  private  databases,  or  data  intermediaries?
[X]  Yes     [  ]  No
If  yes,  specify  the  modality(ies)  of  the  content  covered  by  the  datasets  concerned:
[X]  Text     [X]  Image     [X]  Video     [X]  Audio     [  ]  Other
If  publicly  known,  list  private  datasets  obtained  from  other  third  parties:
Global  Regulation;  TAUS  Translation  Memory;  Scale  HLE;  HackerRank  Coding;  Granary;  SIFT-50M;  Mercor  (agentic  terminal-task  data  and  agentic-search  tool-use  data);  Turing  Math  Data  Pack;  Patronus  (dialog  memory  data);  "Droid  Harness  Pivot"  data;  "Chemistry  Ether0"   data.
General  description  of  non-publicly  known  private  datasets  obtained  from  third  parties:
Text;  professional  translation    memories;  expert-authored  examination  and  reasoning  items;  competitive-programming  problem  sets  and  solutions;  speech  and  speech-instruction;  and  vendor-produced  agentic  task,  tool-use  and  dialogue  datasets  with  human  annotation.
Additional  comments  (optional)
2.3.  Data  crawled  and  scraped  from  online  sources

Were  crawlers  used  by  the  provider  or  on  behalf  of?
[X]  Yes     [  ]  No

Crawler

Names/Identifier:
 NVIDIA  Crawler  Identifier:  nvidiabot  Additional  third-party  crawler  identifiers
unknown.

Purposes  of  the  crawler(s):  Acquire  large-scale  web  text  for  pre-training;  acquire  permissively  licensed  source  code  for  code  pre-training;  acquire  images  and  image-text  pairs  for  captioning,  OCR,  document  understanding  and  visual  question  answering;  and  acquire  video  and  audiovisual  material  for  captioning,  question  answering  and  temporal  reasoning.
General  description  of  crawler  behaviour  Automated  crawlers  were  used  to  collect  publicly  available  media  from  selected  websites.  The  crawlers  navigated  publicly  accessible  pages,  extracted  relevant  links  and  associated  metadata,  and  collected  eligible  content  for  model  training  and  evaluation.  Each  crawl  was  operated  in  accordance  with  the  rate  limits  set  by  its  respective  source.
Period  of  data  collection  2013–2026
Comprehensive  description  of  the  type  of  content  and  online  sources  crawled:
Images  and  documents:  publicly  accessible  web  images,  documents  and  PDFs  used  for  captioning,  OCR  and  document/visual  question  answering.   Video  and  audio:  publicly  accessible  video  and  audiovisual  material  used  for  captioning,  question  answering  and  reasoning.
Type  of  modality  covered:  [X]  Text     [X]  Image     [X]  Video     [X]  Audio     [  ]  Other
Summary  of  the  most  relevant  domain  names  crawled:
github.com;  youtube.com
Additional  comments  (optional)
2.4.  User  data

Was  data  from  user  interactions  with  the  AI  model  (e.g.  user  input  and  prompts)  used  to  train  the  model?
[  ]  Yes     [X]  No
Was  data  collected  from  user  interactions  with  the  provider's  other  services  or  products  used  to  train  the  model?
[  ]  Yes     [X]  No
If  yes,  provide  a  general  description  of  the  provider's  services  or  products  that  were  used  to  collect  the  user  data:
N/A
Type  of  modality  covered:  N/A
Additional  comments  (optional)
2.5.  Synthetic  data

Was  synthetic  AI-generated  data  created  by  the  provider  or  on  their  behalf  to  train  the  model?
[X]  Yes     [  ]  No
If  yes,  modality  of  the  synthetic  data:  [X]  Text     [X]  Image     [X]  Video     [X]  Audio     [  ]  Other
If  yes,  specify  the  general-purpose  AI  model(s)  used  to  generate  the  synthetic  data  if  available  on  the  market:
Alibaba  /  Qwen:  Qwen2.5-0.5B-Instruct;  Qwen2.5-14B-Instruct;  Qwen2.5-32B-Instruct;  Qwen2.5-72B;  Qwen2.5-72B-Instruct;  Qwen2.5-Math-7B;  Qwen2.5-Math-72B;  Qwen2.5-VL-32B-Instruct;  Qwen2.5-VL-72B-Instruct;  Qwen2-VL-72B;  Qwen3-8B;  Qwen3-30B-A3B  (and  Instruct-2507  /  Thinking-2507);  Qwen3-32B;  Qwen3-235B-A22B  (and  Instruct-2507  /  Thinking-2507);

Qwen3-Next-80B-A3B  (Instruct  /  Thinking);  Qwen3-Coder-480B-A35B-Instruct;  Qwen3-VL-30B-A3B  (Instruct  /  Thinking-FP8);  Qwen3-VL-32B-Instruct;  Qwen3-VL-235B-A22B  (Instruct  /  Thinking,  incl.  FP8);  Qwen3-Omni-30B-A3B-Instruct;  Qwen3-omni-captioner;  Qwen3.5-122B-A10B;  Qwen3.5-397B-A17B;  QwQ-32B  DeepSeek:  DeepSeek-V3;  V3-0324;  V3.2;  V3.2-Speciale;  V4-Pro;  R1;  R1-0528;  R1-Distill-Qwen-7B;  R1-Distill-Qwen-32B;  DeepSeek  OCR  OpenAI:  gpt-oss-20b;  gpt-oss-120b  Google:  Gemma-2-2b-it;  Gemma-3-4b-it;  Gemma-3-27b-it;  Gemini  3  Mistral:  Mixtral-8x7B-v0.1;  Mixtral-8x22B-v0.1;  Mixtral-8x22B-Instruct-v0.1;  Mistral-7B-Instruct-v0.3;  Mistral-Small-3.1-24B-Instruct-2503;  Mistral-Small-3.2-24B-Instruct-2506;  Mistral-NeMo-12B-Instruct;  Ministral-3-14B-Reasoning-2512  Microsoft:  phi-4;  Phi-4-mini-instruct;  Phi-3-small-8k-instruct;  Phi-3-medium-4k-instruct;  Florence-2-large  Zhipu  /  Z.ai:  GLM-4.1V-9B-Thinking;  GLM-4.6;  GLM-4.7  (incl.  Flash  and  FP8);  GLM-5  Moonshot:  Kimi-K2-Instruct;  Kimi-K2-Thinking;  Kimi-K2.5  MiniMax:  MiniMax-M1-40k;  MiniMax-M1-80k;  MiniMax-M2   Other:  Goedel-Prover-V2-32B;  Ring-1T;  SwinDocSegmenter

Information  about  other  AI  models,  including  provider's  own  AI  model(s)  not  available  on  the  market,  used  to  generate  synthetic  data  to  train  the  model  to  which  this  Summary  applies:
NVIDIA-Nemotron-3-Ultra-Teacher-STEM  NVIDIA-Nemotron-3-Ultra-Teacher-General-Reasoning  NVIDIA-Nemotron-3-Ultra-Teacher-Office-Work  NVIDIA-Nemotron-3-Ultra-Teacher-Terminal-Use  NVIDIA-Nemotron-3-Ultra-Teacher-Search  NVIDIA-Nemotron-3-Ultra-Teacher-General-Agentic  NVIDIA-Nemotron-3-Ultra-Teacher-Conversational-Tool-Use  NVIDIA-Nemotron-3-Ultra-Teacher-Instruction-Following  NVIDIA-Nemotron-3-Ultra-Teacher-Competitive-Coding  NVIDIA-Nemotron-3-Ultra-Teacher-Chat  NVIDIA-Nemotron-3-Ultra-Teacher-Agentic-Safety    The  teacher  models  were  trained  using  the  same  broad,  multidisciplinary  data  pool  described  above,  with  tailored  data  selection,  mixture  weighting,  synthetic-data  generation,  and  training  recipes  designed  to  emphasize  the  capabilities  associated  with  each  domain,  including  STEM,  reasoning,  office  work,  terminal  and  search  use,  agentic  and  conversational  tool  use,  instruction  following,  competitive  coding,  chat,  and  agentic  safety.
Additional  comments  (optional)
2.6.  Other  sources  of  data

Have  data  sources  other  than  those  described  in  Sections  2.1  to  2.5  been  used  to  train  the  model?
[X]  Yes     [  ]  No
If  yes,  provide  a  narrative  description  of  these  data  sources  and  the  data:
NVIDIA-created  and  NVIDIA-generated  reasoning  and  puzzle  environments,  and  machine-translated  corpora;  internal  competitive-coding,  long-context  and  software-engineering  reinforcement-learning  environments;  internal  CUDA  kernel  task  data;  internal  tool-use  and  agentic  interaction  data;  hand-written  prompt  sets  and  model-identity  data;  and  internal  safety  datasets.
Additional  comments  (optional)
3.  Data  processing  aspects
3.1.  Respect  of  reservation  of  rights  from  text  and  data  mining
exception

or

limitation

Are  you  a  Signatory  to  the  Code  of  Practice  for  general-purpose  AI  models  that  includes  commitments  to  respect  reservations  of  rights  from  the  TDM  exception  or  limitation?
[  ]  Yes     [X]  No
Describe  the  measures  implemented  before  model  training  to  respect  reservations  of  rights  from  the  TDM
NVIDIA  implements  measures  to  respect  EU  rights  reservations  relevant  to  text  and  data  mining  by:  (1)  respecting  robots.txt

exception  or  limitation  before  and  during  data  collection,  including  the  opt-out  protocols  and  solutions  honoured  by  the  provider  or,  as  applicable,  by  third  parties  from  which  datasets  have  been  obtained:
directives  at  the  domains  accessed  where  those  directives  constituted  a  reservation  of  rights  under  Article  4(3)  of  Directive  (EU)  2019/790;  and  (2)  filtering  datasets  on  any  actionable  metadata  identifiers  provided  by  rightsholders.
Additional  comments  (optional)
3.2.  Removal  of  illegal  content

General  description  of  measures  taken  Training  datasets  were  passed  through  layered  automated  and  human  safeguards  designed  to  reduce  the  presence  of  illegal,  harmful  or  policy-violating  content,  including  child  sexual  abuse  material  (CSAM),  non-consensual  intimate  imagery  (NCII),  sexual  content  involving  minors,  weapons  and  criminal-planning  content,  harassment,  hate  speech,  profanity,  threats  and  incitement  to  violence,  self-harm  and  suicide-related  content,  and  graphic  violence.   A  guard  model  fine-tuned  on  the  Nemotron  Content  Safety  Dataset  v2  was  used  to  exclude  potentially  illegal  or  harmful  content  from  training  and  to  quality-filter  synthetically  generated  data.  For  multimodal  data,  two  independent  safety  judge  models  were  applied  to  flag  and  remove  harmful  samples.   Image-based  datasets  were  scanned  against  known  CSAM,  with  the  inventory  recording  passed  compliance  scans  or  reuse  of  prior  scans  for  identical  source  images  where  applicable.  Safety-alignment

datasets,  rule-based  filtering,  and  model-based  filtering  were  also  used.   Datasets  were  subject  to  content-safety  checks  appropriate  to  their  source  and  modality.  Web-derived  image-text  data  underwent  automated  NSFW  screening  designed  to  identify  explicit  imagery  and  profane  or  inappropriate  accompanying  text,  with  flagged  samples  removed  before  using  the  data  for  training.   Synthetic  and  simulation-generated  data  are  evaluated  through  internal  validation  before  inclusion.  Benchmark  evaluations  and  red-team  testing  are  applied  post-training  to  surface  remaining  safety  gaps  across  world  generation,  reasoning,  audio,  and  action  tasks.  Ongoing  monitoring  and  dataset  review  continue  post-release.
3.3.  Other  information  (optional)

Other  relevant  information  about  data  processing  (optional)