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

Pleias_Baguettotron_2026_09_08 — capture 20260912T062204Z

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 74c5db3df4962d252f5a6cc8735eabc1cf1e3bed570b1178bc05265f82a573e5.pdf must equal the hash above (the filename IS the expected hash); ots verify 74c5db3df4962d252f5a6cc8735eabc1cf1e3bed570b1178bc05265f82a573e5.pdf.20260912T062204Z.ots -f 74c5db3df4962d252f5a6cc8735eabc1cf1e3bed570b1178bc05265f82a573e5.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).

All summaries Pleias 1.0 Baguettotron and Monad
Public Summary of Training Content for General-Purpose
AI models: Pleias-Baguettotron, Pleias-Monad
V ersion of the Summary: V ersion 1.0 — first publication of this Summary . No
previous versions.
Last update: 04/08/2026
General information
1. General information
1.1. Provider identification
Provider name and contact details:
PLEIAS, société par actions simplifiée (SAS), share capital
EUR 5,000.
Registered with the Paris Trade and Companies Register
(RCS Paris) under number 982 899 023. Intra-EU V A T
number FR95 982 899 023.
Registered office: 91 rue des Maraîchers, 75020 Paris,
France.
General contact: contact@pleias.fr
Contact point for rightsholders and for questions
concerning this Summary: contact@pleias.fr
W ebsite: https://pleias.ai
Model repositories: https://huggingface.co/PleIAs
Authorised representative name and contact details:
Not applicable. PLEIAS is established in the Union (France).
Article 54 AI Act, which concerns providers established in
third countries, does not apply .
1.2. Model identification
V ersioned model name(s): This Summary covers the following two model versions.
Their training content is identical. They are reported
together in a single Summary in accordance with point (30)
of the Commission Explanatory Notice to the T emplate,
under which the same Summary may be used for different
models or model versions where the content of their
respective Summaries is identical. The model-specific
figures given below — parameter counts and the number
of passes made over the corpus — are disclosed voluntarily
and do not affect the content of the training data, which is
the same in each case.
• PleIAs/Baguettotron — 320,956,992 parameters, 80
layers — https://huggingface.co/PleIAs/Baguettotron
• PleIAs/Monad — 56,656,128 parameters, 64 layers —
https://huggingface.co/PleIAs/Monad
Baguettotron and Monad Training Content | Pleias https://pleias.ai/training-content/baguettotron-monad
1 of 15 08/09/2026, 17:08
Both are decoder-only transformer text-generation models
released under the Apache 2.0 licence, and both were
trained exclusively on the same dataset, SYNTH. The model
card for each is published at the URL given above.
The two models differ in size, in context length (4,096
tokens for Baguettotron, 2,048 for Monad), in tokenizer
(Baguettotron uses the PLEIAS byte-pair-encoding
tokenizer with a vocabulary of 65,536 entries, of which
approximately 45 are reserved as control tokens for
reasoning; Monad uses a purpose-built English-only
tokenizer with a vocabulary of 8,192 entries trained on
SYNTH itself), and in the languages they are documented as
supporting.
Model dependencies:
None. Both models were pre-trained from scratch by
PLEIAS. Neither is a modification, fine-tune, distillation or
other derivative of a pre-existing general-purpose AI model,
and neither declares a base model.
Neither model underwent a separate supervised fine-
tuning, instruction-tuning or reinforcement-learning stage.
Instruction-following and reasoning behaviour was
acquired during pre-training, because the training corpus
consists of instruction and reasoning material throughout.
The training data was itself produced with the assistance of
third-party general-purpose AI models; those models are
identified in Section 2.5. They are generators of the training
data, not components of, or dependencies for , the models
covered by this Summary .
Date  of  placement  of  the  model  on  the  Union
market:
• Baguettotron — repository created and weights
published 10 November 2025; model card published 10
November 2025.
• Monad — repository created and weights published 10
November 2025; model card published 10 November
2025.
Both models were publicly announced on 10 November
2025. Neither has been further trained since its release, and
no new version has been placed on the market. A revision
to Baguettotron's tokenizer files on 27 April 2026 did not
alter the model weights or the training data.
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
Select the modalities present in the training
data, to the extent that they are identifiable
Training data size
For each selected modality , select the range
within which the estimated total training
data size for that modality falls. Dynamic
datasets may be excluded from the
estimation.
Types of content
For each selected modality , provide a general
description of the type of content that has
been included in the training data.
☒ Text
☐ Less than 1 billion tokens
☒ 1billion to 10 trillions tokens
☐ More than 10 trillions tokens
Precise figures, disclosed
voluntarily:
• Training dataset: 79,648,272 text
samples, comprising over 41
Synthetic instructional and reasoning
text, generated in its entirety by AI
models from a fixed set of openly
licensed encyclopaedic seed
documents. No text was taken from
the open web, and no naturally
occurring text other than the seed
Baguettotron and Monad Training Content | Pleias https://pleias.ai/training-content/baguettotron-monad
2 of 15 08/09/2026, 17:08
billion words, equivalent to
approximately 75 billion tokens
with the PLEIAS tokenizer .
• Tokens processed per model:
approximately 200 billion for
each of Baguettotron and
Monad, as stated on the model
cards. For Baguettotron this
corresponds to between two and
three passes over the dataset.
Monad was trained with a
different tokenizer , so the
dataset's tokenised length — and
therefore the number of passes
— differs.
passages quoted within the generated
material is present.
Distribution of the dataset by exercise
type, measured over the released
dataset:
• memorisation — question-and-
answer pairs with reasoning traces,
grounded in an encyclopaedic
passage — 90.6%;
• multiple-choice questions with
distractors — 2.1%;
• constrained writing exercises —
1.7%;
• mathematical exercises with
symbolic solutions — 1.6%;
• retrieval-augmented-generation
exercises with cited sources —
1.0%;
• mathematical multiple-choice
questions — 1.0%;
• creative writing, including
lipograms, layout poems and style
or persona exercises — 0.8%;
• editing exercises, covering
translation, structured extraction,
orthographic and factual correction
and style reformulation — 0.8%;
• practical-knowledge exercises based
on cooking recipes — 0.5%.
Each record carries the question, the
seed passage from which it was
derived, the URL and licence of that
seed, the constraints applied, the
generated reasoning trace and the
generated answer .
Source code was deliberately excluded
from the dataset. No image, audio or
video content was used. No press
publications, fiction or other material
under active copyright protection
were used.
☐ Image
☐ Less than 1 million images
☐ 1Million to1 billion images
☐ More than 1 billion images
Examples of possible types of content include
photography , visual art works, infographics,
social media images, logos, brands.
☐ Audio[1]
☐ Less than 10 000 hours
☐ 10 000 to1 million hours
☐ More than 1 million hours
Examples of possible types of content include
musical compositions and recordings,
audiobooks, radio shows and podcasts,
private audio communication.
☐ Video
☐ Less than 10 000 hours
☐ 10 000 to1 million hours
☐ More than 1 million hours
Examples of possible types of content include
music videos, films, TV programmes,
performances, video games, video clips,
journalistic videos, social media videos.
☐ Other Not applicable. The training data
comprises text only .
Latest date of data acquisition/collection for model
training:
11/2025.
The dataset was generated and frozen before pre-training
Baguettotron and Monad Training Content | Pleias https://pleias.ai/training-content/baguettotron-monad
3 of 15 08/09/2026, 17:08
began. The underlying encyclopaedic seed material was
taken from a Wikipedia snapshot preceding that date.
Neither model is continuously trained on new or dynamic
data after that date, and neither has been further trained
since release.
Description of the linguistic characteristics of the
overall training data:
The dataset is multilingual, with approximately 20% of
samples in languages other than English.
Distribution by language, measured over the released
dataset:
English 80.9%; German 3.16%; Spanish 3.16%; French
3.15%; Polish 3.15%; Italian 3.14%; Dutch 1.60%; Latin
1.60%.
The non-English languages were selected from those best
represented in PLEIAS's earlier Common Corpus dataset.
Reasoning traces are written in English even where the
question and answer are in another language; a question
may be quoted verbatim in its original language within an
English trace.
Of the seven non-English languages present in material
quantity , six are official languages of the European Union
(German, Spanish, French, Polish, Italian, Dutch) and one is
a historical European language (Latin).
Other relevant characteristics of the overall training
data:
The dataset is derived from a deliberately small and fixed
seed corpus of 62,555 documents, each amplified at least
one hundred times, so that the factual knowledge the
models can acquire is bounded by , and traceable to, an
enumerable set of source documents. The seed corpus is
composed of 58,698 Wikipedia articles — 50,000 from the
community-curated "Vital articles" lists (levels 1 to 5) and
8,698 further articles added to reinforce coverage of law ,
medicine and chemistry — together with 3,727 Wikibooks
pages on cooking and practical knowledge and 130
documents written internally by PLEIAS. The 130 internal
documents were amplified approximately ten thousand
times.
PLEIAS states expressly in its published documentation that
the seed selection carries the structural biases of Wikipedia
contribution and editing, and that the selection was made
from the perspective of W estern European and United
States culture. Coverage of knowledge specific to other
regions is correspondingly limited.
Because the dataset contains encyclopaedic information
about well-known historical and public figures and no
naturally occurring personal data, PLEIAS's published
documentation states that no personal-data curation was
required. See Section 3.2.
The dataset is published in full at https://huggingface.co/
datasets/PleIAs/SYNTH under the CC-BY-4.0 licence, so that
the entire training content of these two models is publicly
inspectable, record by record. Each record additionally
reproduces its encyclopaedic seed passage verbatim in a
dedicated field, together with that passage's own URL and
licence; those seed passages remain subject to their own
licence, which is CC-BY-SA 4.0 for 99.46% of records and CC0
for 0.54%.
Baguettotron and Monad Training Content | Pleias https://pleias.ai/training-content/baguettotron-monad
4 of 15 08/09/2026, 17:08
Additional comments (optional):
Token counts are computed with the PLEIAS tokenizer , a
byte-pair-encoding tokenizer with a vocabulary of 65,536
entries. Monad was trained with a different, purpose-built
tokenizer of 8,192 entries; the dataset is identical in both
cases and only its tokenised length differs. Because
tokenizer vocabularies differ between providers, these
token counts are not directly comparable with figures
reported elsewhere.
The dataset occupies approximately 240 GB in its published
Parquet form.
Baguettotron's tokenizer reserves approximately 50 control
tokens used within reasoning traces: markers for logical,
epistemic and verification steps, and simulated entropy
annotations. These are training-time annotations present in
the data; they do not correspond to any inference-time
mechanism.
2 .  L i s t  o f  d a ta  s o u r c e s
1. List of data sources
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
This Section requires information about datasets that were used to train the model and which have been compiled by a third
party , are made available publicly for free, and are readily downloadable as a whole or in predefined chunks, such as datasets
and collections available on public repositories and online platforms, specialised websites, or snapshots of common crawl.
The public availability of the datasets for free does not mean that the content at issue is necessarily free of rights since it may
be subject to licensing arrangements or conditions of use (e.g., certain free and/or open licenses may determine the scope of the
uses, including prohibiting uses relating to model training).
A dataset is considered to be “large” if the total data size for any one of the modalities contained in the dataset exceeds 3% of
the size of all publicly available datasets for that modality used for training. The size of the dataset should be based on its size
after pre-processing (for example filtering), and without splitting the dataset to prevent reporting circumvention.
Have you used publicly available datasets to train the
model?   ☒ Y es     ☐ 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 dataset actually used for training, SYNTH, is synthetic and is therefore reported in
Section 2.5, in accordance with the instruction in Section 2 of the T emplate that synthetic
data be reported in Section 2.5 even where it could also be assigned to another category .
This Section reports the publicly available datasets that were used as the seed material
from which the synthetic data was generated.
1. Structured Wikipedia (Wikimedia Enterprise) — https://huggingface.co/datasets/
wikimedia/structured-wikipedia
Machine-readable dumps of Wikipedia articles published by the Wikimedia Foundation
Baguettotron and Monad Training Content | Pleias https://pleias.ai/training-content/baguettotron-monad
5 of 15 08/09/2026, 17:08
through its Wikimedia Enterprise service. PLEIAS used 58,698 articles from this dataset:
50,000 from the community-curated "Vital articles" lists at levels 1 to 5 (https://
en. wikipedia.org/wiki/Wikipedia:Vital_articles/Level/5) and 8,698 further articles selected
by category-tree and Wikidata-graph expansion to reinforce coverage of law , medicine and
chemistry . Content is licensed CC-BY-SA 4.0. This is the largest seed source and accounts for
the substantial majority of the seed material.
2. Wikibooks — https://www . wikibooks.org
3,727 pages on cooking and practical knowledge, obtained through the official Wikimedia
API, licensed CC-BY-SA 4.0. This category was included because practical and procedural
knowledge is under-represented in encyclopaedic articles.
3. Formalised mathematics templates, mostly from the Kimina dataset - https://
huggingface.co/datasets/AI-MO/Kimina-Prover-Promptset — approximately 3,000
formalised mathematical exercises, of which most came from the Kimina dataset, used as
templates for the mathematical portion of the dataset. V ariable values were randomised
and the symbolic solutions recomputed, so that the released mathematical exercises are
generated rather than reproduced.
Approximate dates of data collection: the seed material was retrieved in 2025 from the
then-current Wikimedia dumps and API. The licence recorded for each individual seed
record is published in the dataset: CC-BY-SA 4.0 for 99.46% of records and CC0 for 0.54%.
General description of
other publicly available
datasets not listed above:
No other publicly available dataset was used. Common Crawl and comparable general
web-crawl corpora were not used.
Additional comments
(optional):
The seed corpus was deliberately kept small and enumerable. PLEIAS's published
rationale is that bounding the seed set bounds the factual knowledge the models can
memorise, and makes it possible to state exactly which documents a model's parametric
knowledge derives from — a property that is not available for models trained on web-scale
corpora.
Every record generated from an encyclopaedic seed carries the URL and the licence of that
seed, so that the provenance of the individual training example can be traced back to its
source article. The mathematical exercises, approximately 1.6% of records, are generated
from formalised templates rather than from an encyclopaedic passage and carry no seed
URL.
Documentation: dataset card https://huggingface.co/datasets/PleIAs/SYNTH; "It's All
Training: A Fully Synthetic Single-Stage Recipe for LLMs" (Langlais et al., preprint); blog
post "SYNTH: the new data frontier", 10 November 2025, https://pleias.fr/blog/blogsynth-
the-new-data-frontier .
2.2 Private non-publicly available datasets obtained from third parties
This Section requires information about private non-publicly available datasets of third parties that are not publicly available
and not disclosed under Section 2.1. These include:
1) datasets for which transactional commercial licensing agreements were concluded between the provider and the
rightsholders or their representatives, including by collective management organisations and legitimate content
aggregators who have the right to collectively license works on behalf of rightsholders (Section 2.2.1);
2) other private datasets obtained through data intermediaries, non-publicly available databases and datasets of third
parties for which transactional commercial licenses have not been concluded with rightsholders  or their
representatives (Section 2.2.2).
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? ☐ Y es     ☒ No
If yes, specify the modality(ies) of the content covered
by the datasets concerned:
Not applicable. No transactional commercial licensing
agreements were concluded with rightsholders or their
representatives.
Baguettotron and Monad Training Content | Pleias https://pleias.ai/training-content/baguettotron-monad
6 of 15 08/09/2026, 17:08
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?
☐ Y es     ☒ No
If yes, specify the modality(ies) of the content covered
by the datasets concerned: Not applicable.
If publicly known, list private datasets obtained from
other third parties: Not applicable. No private dataset was obtained from any
third party .
General description of non-publicly known private
datasets obtained from third parties Not applicable.
Additional comments (optional):
PLEIAS has not obtained any dataset from a data
intermediary , a private database provider or any other
third party on a non-public basis for the training of these
models.
2.3 Data crawled and scraped from online sources
This Section requires information about crawled, scraped data, or otherwise compiled from online sources directly by the
provider of the model or on their behalf (i.e. excluding publicly available datasets already compiled by third parties and made
available on platforms such as common crawl that are covered under Section 2.1).

W ere crawlers used by the provider or on behalf of?☐ Y es     ☒ No
If yes, specify crawler name(s)/identifier(s): Not applicable. No crawler was used by PLEIAS or on its
behalf for the training of these models.
Purposes of the crawler(s): Not applicable.
General description of crawler behaviour:
Not applicable. No content was crawled or scraped from
online sources for the training of these models.
The seed material was obtained from machine-readable
dumps published by the Wikimedia Foundation through its
Wikimedia Enterprise service and from the official
Wikimedia API. Both are interfaces provided by the
rightsholder for the express purpose of bulk reuse of freely
licensed content. Obtaining data through them does not
involve crawling or scraping, and no link-following, rate-
limited harvesting or access to any site other than those
official interfaces took place.
Period of data collection: Not applicable.
Comprehensive description of the type of content
and online sources crawled:
Not applicable. No content was crawled or scraped from
online sources for the training of these models.
Type of modality covered:  ☐ Text   ☐ Image   ☐ Video  ☐ Audio
 ☐ Other  If so, please specify …
Summary of the most relevant domain names
crawled:
Not applicable. No internet domain was crawled or
scraped.
For completeness, the seed material originates from two
domains, accessed through the rightsholder's own bulk-
distribution interfaces rather than by crawling:
wikipedia.org (via Wikimedia Enterprise) and
wikibooks.org (via the official Wikimedia API).
Additional comments (optional): The absence of any crawling is a deliberate design property
of these models rather than an incidental fact. PLEIAS's
stated purpose in building a fully synthetic training corpus
from a small seed set was to demonstrate that a competitive
Baguettotron and Monad Training Content | Pleias https://pleias.ai/training-content/baguettotron-monad
7 of 15 08/09/2026, 17:08
general-purpose model can be trained without recourse to
web-scraped data.
2.4 User data
This Section requires information about user data collected by all services and products of the provider , including through
mail services, social media platforms, content platforms or interaction with the providers’ AI models and/or systems. This
does not cover data licensed by users based on commercial transactional agreements described in Section 2.2.1., or customer
data to fine-tune models for specific purposes.
W as data from user interactions with the AI model
(e.g. user input and prompts) used to train the
model?
☐ Y es     ☒ No
W as data collected from user interactions with the
provider’s other services or products used to train the
model?
☐ Y es     ☒ No
If yes, provide a general description of the provider’s
services or products that were used to collect the user
data:
Not applicable. No data from user interactions with these
models, and no data collected from any other PLEIAS
service or product, was used to train them.
The models are released as open-weight models under the
Apache 2.0 licence and are downloaded and run by third
parties on their own infrastructure. PLEIAS does not
operate a consumer-facing service that collects prompts,
conversations or other user content for training purposes,
and has never used data of that kind in any training stage of
these models.
Type of modality covered: Not applicable.
Additional comments (optional):
PLEIAS confirms that no interaction data of any kind —
including prompts, conversations, uploaded documents,
feedback signals or telemetry — has been collected for or
used in the training of these models at any stage.
2.5 Synthetic data
This Section requires information about synthetic data created by or on behalf of the provider for training the model directly
on the outputs of another AI model, in particular through model distillation or model alignment (e.g. AI feedback through
reinforcement learning). This does not include the use of AI models to clean or enrich data (e.g. AI-generated metadata to
enrich or modify a dataset, such as creating depth maps or text descriptions of images). In case this concerns publicly
available datasets as described in Section 2.1, these should be reported in that Section of the Template. In case this concerns
synthetic datasets created by third parties on behalf of the provider , these should be reported in this Section of the Template
instead of in Section 2.2.2.
W as synthetic AI-generated data created by the
provider or on their behalf to train the model?   ☒ Y es     ☐ No
If yes, modality of the synthetic data:  ☒ Text   ☐ Image   ☐ Video  ☐ Audio       ☐ Other  If so,
please specify …
If yes, specify the general-purpose AI model(s) used to
generate the synthetic data if available on the
market:
Y es. The entire training corpus of these two models is
synthetic: the SYNTH dataset, comprising 79,648,272
generated samples totalling approximately 75 billion
tokens, published at https://huggingface.co/datasets/PleIAs/
SYNTH under the CC-BY-4.0 licence.
The generation pipeline has two stages. In the first, PLEIAS
fine-tunes small auxiliary models on curated examples. In
the second, those fine-tuned auxiliaries are run at scale
over the seed corpus to produce the dataset. The generation
was therefore performed by PLEIAS's own fine-tuned
derivatives, but those derivatives are based on third-party
general-purpose AI models available on the market, which
are identified below .
Baguettotron and Monad Training Content | Pleias https://pleias.ai/training-content/baguettotron-monad
8 of 15 08/09/2026, 17:08
General-purpose AI models on which the generators
are based, all available on the market:
• Qwen 3 8B, provider Alibaba Cloud (Qwen Team) — the
basis of the task-specific fine-tuned models that
generated approximately 97% of the dataset, covering
memorisation, multiple-choice questions, constrained
writing, retrieval-augmented generation, creative
writing, editing and practical-knowledge exercises, and
the symbolic re-evaluation of the mathematical
exercises. Model documentation: https://huggingface.co/
Qwen/Qwen3-8B
• DeepSeek-Prover (7B), provider DeepSeek — used in a
drafter-and-solver configuration for the mathematical
exercises, approximately 2.5% of the dataset. Model
documentation: https://huggingface.co/deepseek-ai/
DeepSeek-Prover-V2-7B
Links to the Summaries of those models, where published
by their respective providers, may be obtained from the
model documentation linked above.
The identity of the model that generated the answer and
reasoning trace is recorded in a dedicated `model` field of
every record of the published dataset, so that the Qwen 3
8B and DeepSeek-Prover attributions can be verified
directly and per record.
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:
Non-generative models used in the pipeline. A frozen
bge-m3 text-embedding model and a F AISS index were
used to retrieve the nearest-neighbour seed paragraph
accompanying each generated question. These perform
retrieval, not generation, and are not general-purpose AI
models.
PLEIAS's own fine-tuned derivatives, not placed on the
market. The task-specific fine-tunes of Qwen 3 8B, the
DeepSeek-Prover drafter and solver configuration
described above were created by PLEIAS for the sole
purpose of generating this dataset. They have not been
placed on the market. Each was fine-tuned on curated
triplets of seed passage, constraints and target output
drawn from the same openly licensed seed corpus
described in Section 2.1, together with the distilled traces
referred to in the confirmation note above.
Selection of the generating models. PLEIAS's position is
that small fine-tuned generators integrated into structured
pipelines give better control over diversity and factual
grounding than prompting a single large model.
Additional comments (optional): The design feature most relevant to the interests this
Summary is intended to serve is that generation proceeds
by *back-translation*: a question is generated from an
authentic seed passage, so that every training target is
anchored to a verifiable statement in a freely licensed
encyclopaedic source, rather than to the unconstrained
output of a large model. PLEIAS's stated purpose is to limit
the propagation of teacher-model errors and to keep the
factual content of the models traceable.
Approximately 20% of the generated queries deliberately
target a refusal, a correction or a hedge rather than a
confident answer , so that the models learn to decline when
Baguettotron and Monad Training Content | Pleias https://pleias.ai/training-content/baguettotron-monad
9 of 15 08/09/2026, 17:08
a question is ill-posed or unanswerable, as described in the
accompanying technical report.
The generated question, reasoning trace and answer are
paraphrases of the seed passage rather than reproductions
of it. Each published record additionally carries the seed
passage verbatim in a dedicated field, together with the
seed URL and licence, so that the freely licensed source of
each record is directly inspectable.
The dataset was released jointly by PLEIAS and the AI
Alliance, and was designed around a set of open standards
for synthetic data developed with the AI Alliance.
2.6 Other sources of data
This Section requires information about data that does not fall under any of the categories in the previous Sections, for
example data collected from offline sources, self-digitised media (e.g., digitised analog text context, images), datasets labelled
by humans commissioned by the provider , or human generated data through reinforcement learning.
Have data sources other than those described in
Sections 2.1 to 2.5 been used to train the model? ☒ Y es     ☐ No
If yes, provide a narrative description of these data
sources and the data:
Y es, in one limited respect. The seed corpus includes 130
documents written internally by PLEIAS staff, covering
documentation of the models themselves, information
about their training conditions, general information on AI
research, and recent events postdating the Wikipedia
snapshot used for the other seeds. These are human-
authored documents created by the provider , and so fall
outside Sections 2.1 to 2.5.
Although they represent only 130 of the 62,555 seed
documents, these texts were amplified approximately ten
thousand times in the generation process — a substantially
higher amplification factor than the minimum of one
hundred applied to the encyclopaedic seeds — because the
knowledge they carry is not otherwise present in the seed
corpus. They are disclosed here notwithstanding their
small number , because their influence on the models' self-
description and on their statements about recent events is
disproportionate to their size.
The documents are original works authored by PLEIAS.
They contain no third-party content, no personal data of
third parties, and no material obtained from any external
source.
Additional comments (optional):
No other data source outside Sections 2.1 to 2.5 was used. In
particular , no offline or self-digitised material, no dataset
labelled by human annotators commissioned by PLEIAS,
and no human-generated data obtained through
reinforcement learning was used to train these models. The
manual curation applied to the fine-tuning sets of the
auxiliary generators was performed internally on
machine-generated text and introduced no new source of
data.
1. Data processing aspects
3. Data processing aspects
Baguettotron and Monad Training Content | Pleias https://pleias.ai/training-content/baguettotron-monad
10 of 15 08/09/2026, 17:08
3.1. Respect of reservation of rights from text and data mining exception or
limitation
This Section concerns measures implemented by the provider to identify and comply with the reservation of rights from the
text and data mining (TDM) exception or limitation expressed pursuant to Article 4(3) of Directive (EU) 2019/790, as outlined
in the copyright policy put in place by the provider in accordance with Article 53(1)(c) AI Act.
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?
☒ Y es     ☐ 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:
PLEIAS is a signatory to the General-Purpose AI Code of
Practice and applies its Copyright chapter commitments.
For these two models the reservation-of-rights question
arises in a materially narrower form than for models
trained on crawled data, because no content was crawled
or scraped and the entire training corpus was generated
from a fixed set of 62,555 seed documents, of which 62,425
are freely licensed third-party documents and 130 are
original works authored by PLEIAS.
1. No crawling or scraping of the open web. No crawler
was operated by PLEIAS or on its behalf for these models,
and no content was collected by crawling or scraping any
website. Content was obtained exclusively through the
rightsholder's own bulk-distribution channels. It follows
that PLEIAS did not access any content in respect of which a
rightsholder had expressed a reservation of rights pursuant
to Article 4(3) of Directive (EU) 2019/790, whether by
machine-readable means or otherwise.
2. Seed material obtained under a free licence, through
the rightsholder's own distribution channel. The seed
corpus consists of Wikipedia and Wikibooks content,
licensed CC-BY-SA 4.0 (99.46% of records) or CC0 (0.54%),
obtained from machine-readable dumps published by the
Wikimedia Foundation through Wikimedia Enterprise and
from the official Wikimedia API. Both are channels
operated by the rightsholder for the express purpose of
bulk reuse. The reproduction of that content therefore rests
on the licence granted by the rightsholder and not on the
text-and-data-mining exception, so that a reservation of
rights under Article 4(3) has no application to it.
3. Generated content is paraphrase, not
reproduction. The questions, reasoning traces and
answers that constitute the training targets are paraphrases
of the seed passages rather than reproductions of them.
Each published record additionally reproduces its seed
passage verbatim in a dedicated field, so that the source of
the record is inspectable; that passage is freely licensed
encyclopaedic text and travels with its own licence notice.
4. Per-record provenance. Every record generated from
an encyclopaedic seed carries the URL and the licence of
that seed, together with the identity of the model that
generated the answer . A rightsholder can therefore
establish, record by record, what seed material was used
and under what licence. The mathematical exercises,
approximately 1.6% of records, are generated from
formalised templates and carry no seed URL.
Baguettotron and Monad Training Content | Pleias https://pleias.ai/training-content/baguettotron-monad
11 of 15 08/09/2026, 17:08
5. Reuse of third-party model outputs. The general-
purpose AI models used to generate the dataset were
selected on the basis that their licences permit the reuse of
their outputs, including for training purposes, so that no
restriction attaching to a third-party model's terms of use
was disregarded.
6. Complaints and removal. Rightsholders who consider
that material has been included in error may write to
contact@pleias.fr . PLEIAS examines such notifications and
acts on those that are well founded, including by removing
material from the published dataset.
Additional comments (optional):
PLEIAS's copyright policy under Article 53(1)(c) AI Act is
published at pleias.ai/training-content
The licensing analysis underlying the corpus is
documented in arXiv:2506.01732, and in particular in its
Appendix E on the verification of public-domain status.
PLEIAS notes that the same Appendix records that it did not
attempt to establish public-domain status on the basis of
non-renewal of United States copyright; where United
States material published after 1929 is present, its public-
domain status rests on the determination made by the
digitising institution.
3.2 Removal of illegal content
This Section concerns measures taken to avoid or remove illegal content under Union law from the training data (such as
blacklists, keywords, and model-based classifiers), without requiring disclosure of specific details about the provider’s internal
business practices or trade secrets. Such measures are advisable if the training data is likely to include illegal or unlawful
content under Union law , in particular child sexual abuse material and terrorist content and the non-authorised use of
material protected by intellectual property rights. Such measures do not include data selection practices, for example to
increase the capability of the model.
General description of measures taken: Composition of the training data as the primary
measure. The training corpus was generated from 62,555
seed documents drawn almost entirely from community-
curated Wikipedia and Wikibooks articles. No content was
taken from the open web, from social media, from forums
or from any user-generated-content platform. The
categories of illegal content associated with large-scale web
scraping — child sexual abuse material, terrorist content
and unauthorised reproductions of protected works — are
therefore excluded at source rather than filtered out after
the fact. PLEIAS's published documentation states that the
systematic grounding of every record in Wikipedia gives
the dataset a very low risk of toxic or otherwise
problematic content.
Personal data. The seed corpus consists of encyclopaedic
material concerning well-known historical and public
figures, drawn from articles that Wikipedia's own editorial
and biographies-of-living-persons policies already govern.
PLEIAS's published documentation states expressly that, for
this reason, no personal-data curation step was required
for this dataset. No naturally occurring personal data —
such as would arise from scraped social media, forums or
web pages — is present, because no such source was used.
The 130 internal seed documents were authored by PLEIAS
and contain no third-party personal data.
Independent quality assessment. As reported in the
accompanying technical report, the dataset was scored
Baguettotron and Monad Training Content | Pleias https://pleias.ai/training-content/baguettotron-monad
12 of 15 08/09/2026, 17:08
alongside a range of open pre-training corpora (including
Nemotron-CC, Fine W eb-2, FinePDFs, Fine Wiki, HPL T-4 and
Common Corpus) using the Propella-1 multilingual
document annotator , a model developed by a third party .
On the integrity and safety axes of that assessment SYNTH
scores at the ceiling, matching Wikipedia-derived corpora
and above web-cleaned corpora.
Exclusion of source code. Code data was deliberately
excluded from the dataset, removing a category of content
that carries distinct licensing and security risks.
Filtering of generations. Generations were filtered during
production to remove defective outputs. The generating
models — Qwen 3 8B, DeepSeek-Prover and Gemma 3 12B
— carry their own provider-level safety mitigations.
Limitations stated in good faith. PLEIAS states in its
published documentation that the seed selection carries the
structural biases of Wikipedia contribution and editing and
reflects a W estern European and United States perspective;
and that at these parameter counts, and particularly for
Monad at 56 million parameters, factual hallucination is to
be expected. These models were released as base models
and have not undergone a separate safety-alignment stage.
Notifications concerning illegal content in the published
dataset may be sent to contact@pleias.fr .
3.3. Other information (optional)
Other relevant information about data processing
(optional):
Scope of this Summary . In accordance with point (30) of
the Commission Explanatory Notice, this single Summary
covers two model versions whose training content is
identical.
Related Summaries. PLEIAS publishes two further
Summaries covering its other model families: one for the
Pleias 1.0 Preview family (Pleias-350m-Preview , Pleias-1.2b-
Preview , Pleias-3b-Preview) and one for the Pleias-RAG
family (Pleias-RAG-350M, Pleias-RAG-1B). Those models
were trained on a different corpus and are not covered
here.
Openness and verifiability . The entire training corpus of
these two models is published, record by record, under the
CC-BY-4.0 licence. Each record carries the seed URL, the
seed licence, the generating model, the constraints applied,
the generated reasoning trace and the generated answer .
Every factual assertion in this Summary concerning the
composition, provenance, language distribution and
generation of the training data can therefore be verified
directly against the published dataset rather than taken on
trust:
• dataset — https://huggingface.co/datasets/PleIAs/SYNTH
• models, under Apache 2.0 — https://huggingface.co/
PleIAs/Baguettotron and https://huggingface.co/PleIAs/
Monad
• training framework — Nanotron, https://github.com/
huggingface/nanotron
Primary references.
• Langlais, Delobelle, Detrois et al., "It's All Training: A Fully
Baguettotron and Monad Training Content | Pleias https://pleias.ai/training-content/baguettotron-monad
13 of 15 08/09/2026, 17:08
Synthetic Single-Stage Recipe for LLMs" (preprint).
• "SYNTH: the new data frontier", PLEIAS blog, 10
November 2025, https://pleias.fr/blog/blogsynth-the-new-
data-frontier .
Compute, disclosed voluntarily . Both models were
trained on 16 NVIDIA H100 GPUs at the Jean Zay
supercomputer (GENCI/IDRIS, compute plan A0191016886)
using the Nanotron framework. Monad's full pre-training
took a little under six hours. The final training runs for both
models together represent fewer than 1,000 H100-hours;
the project as a whole, including the generation of the
dataset and all preparatory experiments, amounted to
approximately 20,000 H100-hours.
Governance. The dataset was released jointly by PLEIAS
and the AI Alliance, and was designed around a set of open
standards for synthetic data developed in collaboration
with the AI Alliance.
Updating. This Summary will be updated in accordance
with point (29) of the Commission Explanatory Notice if
these models are further trained on additional data. No
such further training has taken place to date.
[1] Excluding audio that is part of video, as this should be reported under the “video” modality instead. Furthermore, the Commission understands the modality of
‘audio’ to include ‘speech’.
Baguettotron and Monad Training Content | Pleias https://pleias.ai/training-content/baguettotron-monad
14 of 15 08/09/2026, 17:08
Contact
Station F, 5 Parv. Alan Turing 75013, Paris
contact@pleias.fr
LinkedIn
GitHub
HuggingFace
©2026 Pleias. All rights reserved
Baguettotron and Monad Training Content | Pleias https://pleias.ai/training-content/baguettotron-monad
15 of 15 08/09/2026, 17:08