GPAI Ledger › Apertus v1.5 (Swiss AI Initiative) › Capture 17 Aug 2026

Apertus v1.5 — capture 20260817T080512Z

Verify: sha256sum 09275f31dffa639f993c8727ccc10f950e1e4fa6c6a15be9d9c44989577e9fd4.pdf must equal the hash above (the filename IS the expected hash); ots verify 09275f31dffa639f993c8727ccc10f950e1e4fa6c6a15be9d9c44989577e9fd4.pdf.ots -f 09275f31dffa639f993c8727ccc10f950e1e4fa6c6a15be9d9c44989577e9fd4.pdf (opentimestamps.org) proves the capture time (fresh proofs report 'pending' until bitcoin-anchored, typically within a day).

Extracted text

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

Public  Summary  of  Training  Content  for  General-Purpose
AI

models

 This  template  is  provided  by  the  European  Commission  and  required  to  be  filled  in  by  providers  of  general-purpose  AI  models  prior  to  their  placing  on  the  Union  market  in  order  to  comply  with  their  obligation  under  Article  53  (1)(d)  of  Regulation  (EU)  2024/1689  (AI  Act).   For  more  information  and  guidance  see  Commission’s  Explanatory  Notice  and  Template  for  the  Public  Summary  of  Training  Content  for  general-purpose  AI  models  |  Shaping  Europe’s  digital  future.
 Version  of  the  Summary:   v1.5  Last  update:   17.7.2026   General  information
1.  General  information  1.1.  Provider  identification
Provider  name  and  contact  details:
Swiss  National  AI  Institute  —  https://www.swiss-ai.org/ —  llm-requests@swiss-ai.org
Authorised  representative  name  and  contact  details:
Only  applicable  if  the  provider  is  established  outside  the  Union  (see  Article  54  AI  Act).
 1.2.  Model  identification
Versioned  model  name(s):
Apertus-1.5-8B  Apertus-1.5-70B  Apertus-1.5-8B-Instruct  Apertus-1.5-70B-Instruct  https://huggingface.co/collections/swiss-ai/apertus-1.5  Model  dependencies:  N/A  Date  of  placement  of  the  model  on  the  Union  market:  July  14,  2026
1.3  Modalities,  overall  training  data  size  and  other  characteristics  This  Section  requires  general  information  about  the  overall  training  data  after  pre-processing  and  before  the  training  of  the  model.    Modality  Training  data  size  Types  of  content
 ☒  Text
☐  Less  than  1  billion  tokens  ☐  1billion  to  10  trillions  tokens  ☒  More  than  10  trillions  tokens   Alternatively,  specify  the  approximate  size  in  a  different  measurement  unit:   17  trillion  tokens
 Public  text-only  datasets  derived  mainly  from  web  documents  written  in  over  1000  languages,  while  making  significant  efforts  were  made  to  respect  consent  by  right  holders.  We  ensure  training  data  is  fully  transparent  and  reproducible,  so  as  to  make  Apertus  an  open-data  and  open-weights  model.  Data  is  filtered  for  robots.txt  compliance  and  personal  identification  information  when  applicable.
1

☒  Image  ☐  Less  than  1  million  images  ☒  1Million  to1  billion  images  ☐  More  than  1  billion  images
Public  image-only  and  image-text-pairs,  gathered  from  permissive  datasets,  themselves  derived  mainly  from  open  image  repositories  and  web  data.  Data  is  filtered  for  robots  compliance  and  personal  identification  information  when  applicable.
☒  Audio  ☐  Less  than  10  000  hours  ☒  10  000  to  1  million  hours  ☐  More  than  1  million  hours
Public  audio-only  and  audio-transcripts-pairs,  gathered  from  permissive  datasets,  themselves  derived  mainly  from  open  initiatives  (such  as  commonvoice)  and  web  data.   ☐  Video  ☐  Less  than  10  000  hours  ☐  10  000  to1  million  hours  ☐  More  than  1  million  hours

☐  Other
Specify  the  modality  and  for  each  one  indicate  approximate  size  and  unit  of  measurement

  Latest  date  of  data  acquisition/collection  for  model  training:
Pre-training  dataset  knowledge  cutoff  is  April  28th,  2026.
Description  of  the  linguistic  characteristics  of  the  overall  training  data:
Languages .  The  pretraining  dataset  includes  more  than  1000  languages  (1782  language-script  pairs),  as  provided  by  the  FineWeb-2 and  FineWeb-2-HQ datasets  respectively.  The  amount  of  data  per  language  reflects  the  natural  frequency  of  web  data  in  each  language,  thus  improving  representation  of  many  communities  with  languages  not  present  in  most  leading  LLMs  yet.   Domain-specificity.  Emphasis  has  been  placed  on  legal  (with  datasets  such  as  Multi-Legal-Pile or  Swiss-caselaw),  medical  (with  datasets  such  as  MedTrinity),  as  well  as  science-related  (finemath)  contents  for  this  release.

Other  relevant  characteristics  of  the  overall  training  data:
Selected  datasets  (or  dataset  subsets)  adhere  to  our  strict  policy  of  full  license  permissiveness  (excluding  sharealike,  non-commercial  and  any  other  restrictive  licenses).  Data  has  been  rigorously  filtered  for  respecting  consent  by  website  owners  (opt-out  for  AI  crawlers,  also  retroactively),  remove  PII,  remove  toxic  content,  and  avoid  verbatim  memorization  during  model  training  (see  below).  Additional  comments  (optional):
The  Apertus  tokenizer  builds  upon  the  Mistral  v3  (tekken),  and  is  used  for  data  size  statistics  (see  also  the  Apertus  technical  report).

2.  List  of  data  sources
2.  List  of  data  sou r c es  This  Section  requires  information  about  specific  sources  of  data  used  to  train  the  general-purpose  AI  model.  In  this  section  “dataset”  should  be  understood  as  a  single,  pre-packaged  collection  of  data.  The  filtering  and  pre-processing  of  data  collected  from  the  same  pre-packaged  collection  should  not  be  considered  a  new  dataset  to  be  disclosed  separately  in  the  sections  below .  If  a  particular  dataset  can  be  assigned  to  more  than  one  of  the  categories  below,  providers  should  select  the  most  relevant  category  and  only  report  the  dataset  in
that

category,

except

in

the

case

of

synthetic

data

(see

Section

2.5).

  2.1.  Publicly  available  datasets
2

Have  you  used  publicly  available  datasets  to  train  the  model?    ☒  Yes      ☐  No If  yes,  specify  the  modality(ies)  of  the  content  covered  by  the  datasets  concerned:
☒  Text    ☒  Image    ☐  Video   ☒  Audio  ☐  Other
If

so,

please

specify…

List  of  large  publicly  available  datasets:
The  following  large  pretraining  datasets  were  not  used  as  is,  but  were  additionally  filtered  for  opt-out  retrospectively,  for  toxicity,  high  quality,  and  other  preprocessing  as  detailed  below.   The  same  versions  and  filtering  techniques  were  used  for  the  datasets  used  both  in  Apertus  v1  and  v1.5
 1
 and  are  detailed  in  our  first  technical  report.  For  datasets  used  only  in  v1.5,  filtering  and  pre-processing  was  very  similar  to  v1.  We  used  the  same  codebase
 2
 and  the  same  list  of  robots.txt.  The  same  PII-removal  filter  was  used.  Slight  overhauls  mainly  concerned  technical  implementation  details.  For  robots-filtered  image  datasets  (MINT-1T,  pixmo-cap),  we  only  excluded  the  sources  that  were  within  our  list  and  had  robots.txt  indications  (i.e.  not  excluding  sources  that  were  not  gathered  by  our  list).    With  respect  to  training  from  a  purely  technical  perspective,  the  most  important  change  is  naturally  the  addition  of  multimodal  data.   In  addition  to  scale,  we  focused  on  data  quality  through  curation,  deduplication,  quality  filtering  and  domain  relevance.  Our  resulting  data-mixture  supports  robust,  general-purpose,  and  specialised  capabilities  while  advancing  open  and  responsible  AI  development.
Text  Datasets   HuggingFaceFW/fineweb-2:  Large-scale,  high-quality  multilingual  web  text  corpus  derived  from  filtered  Common  Crawl  data.  Serves  as  a  primary  general-purpose  pretraining  source  for  LLMs,  with  strong  emphasis  on  diversity  and  reduced  noise.  License:  ODC-BY.  (Subject  to  PII  removal  and  robots.txt  filtering.)
HuggingFaceTB/dclm-edu:  Curated  educational  web  dataset  from  the  DataComp-LM  project.  Provides  high-signal,  knowledge-rich  text  optimised  for  reasoning  and  factual  learning  in  language  models.  License:  CC-BY-4.0.  (PII  removal  and  robots.txt  filtering  applied.)
nvidia/Nemotron-CC-v2.1:  NVIDIA-curated  Common  Crawl  dataset  (v2.1)  featuring  quality  scoring  and  filtering  for  large-scale  LLM  pre-training.  License:  Nvidia’s  custom  data  and  model  license  (permissive,  see  dataset  page  for  terms).
HuggingFaceFW/finePDFs-edu:  High-quality  collection  of  educational,  scientific,  and  technical  PDF  documents  with  extracted  clean  text.  Enhances  long-context  and  domain  knowledge  capabilities.  License:  ODC-BY.  (Filtered  for  quality  and  compliance.)
joelniklaus/Multi_Legal_Pile:  Specialized  multilingual  legal  corpus  aggregating  statutes,  case  law,  contracts,  and  regulatory  texts  from  multiple  jurisdictions.  Strengthens  legal  reasoning  and  domain  adaptation.  License:  only  compliant  (non-SA,  non-NC)  subsets  of  this
2
 Public  version  accessible  at:  https://github.com/swiss-ai/pretrain-data

1
 The  following  datasets  are  concerned:  dclm-edu,  fineweb-2,  finetranslations,  finemath,  MegaMath,  stack-edu.
3

compound  dataset  were  used.  (PII  removal  and  robots.txt  filtering  applied.)
nvidia/Nemotron-Pretraining-Code-v1:  Large-scale,  curated  code  corpus  from  NVIDIA  designed  to  boost  programming,  software  engineering,  and  logical  reasoning  abilities.License:  Nvidia’s  custom  data  and  model  license  (permissive,  see  dataset  page  for  terms).
Audio  Datasets  mozilla/CommonVoice24:  Mozilla’s  crowdsourced  multilingual  speech  corpus  with  validated  transcriptions  across  many  languages  and  accents.  A  cornerstone  dataset  for  inclusive,  robust  automatic  speech  recognition  (ASR)  and  text-to-speech  (TTS).  License:  CC-BY-1.0.
speechcolab/gigaspeech:  Large-scale  English  speech  recognition  corpus  (~10k  hours)  sourced  from  audiobooks,  podcasts,  and  YouTube  with  high-quality  transcriptions.  Supports  general-domain  ASR  and  audio  understanding.  License:  Apache-2.0.
MLCommons/peoples_speech:  One  of  the  largest  publicly  available  multilingual  speech  datasets,  containing  tens  of  thousands  of  hours  of  diverse,  real-world  speech.  License:  CC-BY-SA  /  CC-BY.
facebookresearch/voxpopuli:  Large  multilingual  speech  corpus  extracted  from  European  Parliament  sessions,  covering  numerous  EU  languages  with  aligned  transcripts.  Excellent  for  cross-lingual  and  parliamentary-domain  speech  tasks.  License:  CC-BY-1.0.
k2-fsa/libriheavy:  Massive  clean  English  read-speech  corpus  (tens  of  thousands  of  hours)  built  on  LibriSpeech  audiobooks  with  precise  alignments.  Known  for  high  acoustic  quality  and  utility  in  ASR/TTS  research.  License:  Apache-2.0.
facebook/omnilingual-asr-corpus:  Massive  multilingual  speech  corpus  spanning  a  wide  range  of  languages  and  dialects,  created  to  advance  open  ASR  systems  globally.  License:  CC  BY  4.0.
Image  Datasets  mlfoundations/MINT-1T:  Landmark  large-scale  image-text  dataset  scaled  to  trillions  of  tokens,  specifically  designed  to  dramatically  expand  open-source  multimodal  pretraining  data.  License:  CC-BY-4.0.  (Includes  PII  removal  and  robots.txt  filtering.)
UCSC-VLAA/Recap-DataComp-1B:  Billion-scale  image-text  dataset  featuring  high-quality  recaptions  of  the  original  DataComp-1B  collection.  Optimized  for  vision-language  pretraining  and  detailed  visual  understanding.  License:  CC-BY-4.0.
dclure/laion-aesthetics-12m-umap:  Curated  12-million-image  subset  of  LAION  focused  on  high  aesthetic  quality  (via  CLIP  and  aesthetic  scoring).  Popular  for  training  visually  pleasing  image  understanding  and  generation  models.  License:  MIT.
mvp-lab/LLaVA-OneVision-1.5-Mid-Training-85M:  Large-scale  (85M  samples)  mid-training  dataset  for  vision-language  models,  emphasizing  instruction  tuning  and  multimodal  alignment.  License:  Apache-2.0.
4

DeepGlint-AI/DanQing100M:  Large-scale  Chinese  image-text  pretraining  dataset  (100M  pairs)  supporting  enhanced  multilingual  and  culturally  relevant  visual-language  capabilities.  License:  CC-BY-4.0.
UCSC-VLAA/MedTrinity-25M:  Large  medical  image-text  dataset  (25M  samples)  with  rich  annotations  and  captions.  Key  resource  for  building  specialized  medical  vision-language  understanding  and  diagnostic  assistance  features.  License:  only  compliant  (non-SA,  non-NC)  subsets  of  this  compound  dataset  were  used  (see  full  list  on  the  dataset’s  page).
 General  description  of  other  publicly  available  datasets  not  listed  above:
Other  smaller  publicly  available  and  permissively  licensed  datasets  were  used  for  the  purposes  described  above.  The  exhaustive  list  of  pre-training  datasets  is  accessible  on  our  GitHub.  Generally,  all  republished  datasets  used  for  Apertus  v1.5  will  be  made  publicly  available  in  the  dedicated  collection  on  Hugging  Face.
Additional  comments  (optional):
Training  data  preparation  code  (for  pre-  and  post-training)  is  entirely  reproducible  and  is  or  will  be  made  publicly  available  in  the  following  repositories:  -  Version  1  text  pretraining  data  preparation  pipelines (v1.5  soon  to  come) -  Version  1  text  posttraining  data  preparation  pipelines (v1.5  soon  to  come) -  Version  1.5  multimodal  data  preparation  pipelines
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      ☒  No
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?
☐  Yes      ☒  No
Additional  comments  (optional):  -
2.3  Data  crawled  and  scraped  from  online  sources   Were  crawlers  used  by  the  provider  or  on  behalf  of?
☒  Yes      ☐  No
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      ☒  No
5

Was  data  collected  from  user  interactions  with  the  provider’s  other  services  or  products  used  to  train  the  model?
 ☐  Yes      ☒  No

2.5  Synthetic  data   Was  synthetic  AI-generated  data  created  by  the  provider  or  on  their  behalf  to  train  the  model?
 ☒  Yes      ☐ No

6

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      ☒  No

1.  Data  processing  aspects  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      ☒  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:
  We  respect  standard  machine-readable  opt-out  by  all  websites.  In  addition,  we  remove  data  from  websites  which  have  recently  opted  out  by  specifying  at  least  one  of  the  common  AI  crawlers,  at  the  time  of  January  2025.  Crucially,  we  have  applied  such  removals  also  retroactively  in  all  earlier  crawls  since  2013,  of  each  corresponding  website  present  in  our  datasets.  Pretraining  and  posttraining  datasets  were  additionally  filtered  for  licence  compliance,  and  text  datasets  are  processed  by  PII  removal.
Additional  comments  (optional):
 Our  Acceptable  Use  Policy  of  Apertus  LLM  is  documented  Usage  Policy  statement.  A  summary  of  our  copyright  policy  can  be  found  on  the  model  card  (Legal  aspects),  and  in  our  Code  of  Practice.
 3.2  Removal  of  illegal  content
General  description  of  measures  taken:
Before  training,  we  remove  toxic  documents  from  the  pretraining  corpora.  For  text  data,  we  do  so  by  employing  a  deep  learning  classifier  trained  on  top  of  XLM-Roberta  multilingual  embeddings.  The  classifiers  were  trained  using  the  multilingual  datasets  provided  by  https://github.com/Pleias/toxic-commons .  In  addition  to  toxicity  filtering,  most  datasets  also  were  filtered  by  additional  quality  classifiers,  which  we  make  transparently  available,  and  which  further  help  to  reduce  problematic  content.  Image  dataset  are  deduplicated  and  converted  to  appropriate  formats  and  dimensions,  among  other  common  preprocessing  practices.
7

3.3.  Other  information  (optional)
Other  relevant  information  about  data  processing  (optional):
In  addition  to  the  full  transparency  of  ensuring  all  training  data  of  our  models  is  openly  available  and  reproducible  (Apertus  models  being  open-data  and  open-weights),  we  also  employed  techniques  to  minimize  memorization  or  potentially  remaining  copyrighted  content:  During  training,  we  employ  the  Goldfish  loss  technique ,  which  disables  verbatim  memorization  of  text  sequences  longer  than  50  tokens.  More  precisely,  every  50th  token  (on  average)  of  our  pretraining  data  is  not  provided  a  prediction  target,  i.e.  has  no  loss  function,  and  thus  breaks  any  verbatim  memorization  beyond  that  sequence  length.  We  provide  more  detailed  results  on  the  success  of  this  mitigation  technique  in  the  model’s  technical  report.
8