README.md

---
configs:
  - config_name: global_academic_theses
    data_files:
      - split: train
        path: global_academic_theses/gac_sample_0*.parquet
  - config_name: youtube_commons
    data_files:
      - split: train
        path: youtube_commons/*.parquet
  - config_name: global_patents
    data_files:
      - split: train
        path: global_patents/*.parquet
  - config_name: telecom
    data_files:
      - split: train
        path: telecom/*.parquet
  - config_name: finance
    data_files:
      - split: train
        path: finance/*.parquet
  - config_name: wikidata_seed
    data_files:
      - split: train
        path: wikidata_seed/*.parquet
  - config_name: personas
    data_files:
      - split: train
        path: personas/*.parquet
  - config_name: synthetic_social_media
    data_files:
      - split: train
        path: synthetic_social_media/*.parquet
---

# Multimodal Pretraining
This section covers our large-scale collections at the source and is distributed in its original form (PDF with layout intact, audio attached to its transcript) rather than as text extracted after the fact. The emphasis is on what large-scale web collection misses: academic global production (badly indexed in scientific repositories); patents outside the US; the technical and regulatory archives of telecom and finance. 

These are long, structured, image-dense documents, which makes them a good fit for long-context and long-horizon multimodal training. Samples mirror the structure of the full assets.

## Global academic thesis

We hold the largest collection of academic thesis with currently 3,134,100 documents in PDF at a global scale. 

Our coverage extends significantly beyond western sources, with about half of our holding (1.5M documents) are not even indexed in leading scientific databases like OpenAlex. Typically, this includes 529,088 full-text thesis from India, 566,385 from Indonesia, and 172,260 thesis from Peru, all hardly indexed let alone collected. The collection is uniquely fitting for long horizon multimodal tasks.

The current sample include 1,015 thesis in pdf formats in 26 languages (English 252, Spanish 97, French 90, Hindi 82…) and most scientific disciplines, along with metadata and new transcriptions by a SOTA VLM for OCR within the language. 

The sample has an average span of 171 pages, rich in multimodal content (50,000 incrusted images, 50 per documents, nearly 60% with associated captions). This collection should fit particularly well the demands of long context training or multimodal long horizon tasks.

## Youtube-Commons

Youtube-Commons is the current reference dataset for Youtube videos under creative commons allowing for full commercial reuse. The dataset was noticeably used to train Nvidia's SOTA text to speech model <a href="https://huggingface.co/nvidia/parakeet-tdt-0.6b-v2">Parakeet</a>.

While the public version of Youtube-Commons only includes metadata and transcripts and has not been updated since 2024, our internal asset includes audio attached to transcripts and metadata for 4,363,162 videos (1 million hours). About 40% of the corpus is Non-English, with more than 100,000 samples in Spanish, Hindi, French, Russian, Portuguese or Korean.

We provide an initial sample of 1,136 audio files from YouTube (394 hours) published from 2009 to 2026 in 93 languages and 168 audio files from VoxPopuli/Europarl (84 hours) published from 2009–2020.

Samples use the same structure as the final dataset and are distributed as 12 fully shuffled parquet files with three components:
* Metadata scraped from the Youtube official API with original url, channel information, date of publication.
* New transcripts made with state of the art ASR model, *not* the Youtube transcript (frequently faulty). These allow for full text search of the entire corpus.
* Audio included directly in the parquet file, as is now common expectation for multimodal model training.

## Global Patents

We are currently collecting the largest available asset of non-US patents. It currently includes the complete collection of European patents (EPO) in the original pdf format and we're currently scaling our data collection to Asia.

Patents commonly come with technical diagrams we can communicate in a separate forms (for now, about 20 million images). The diagrams are described in other parts of the text with complex layout and structure, providing unresolved challenges for frontier multimodal domains across a wide range of disciplines and technical expertise.

The current sample includes 5,000 European patents published 2020–2025 (about 833 per year), distributed as 5 fully shuffled parquet files of 1,000 documents each. Following the same structure as our other multimodal samples, the original PDF is embedded directly in the parquet rather than shipped separately, alongside the full text and metadata.

The sample spans 143,307 pages (29 per patent on average, up to 819) and 57.7 million words, across all 8 IPC sections and 437 distinct technology classes, from 3,042 distinct assignees including Samsung, Microsoft, BASF, Boeing and Nokia.

# Specialized per-domain corpora

## Telecom

In partnership with GSMA we collected <a href="https://huggingface.co/datasets/GSMA/Telco-Common-Corpus">Telco-Common Corpus</a>, currently the largest dataset of telecom data. It has been incorporated into the pretraining of <a href="https://www.gsmaintelligence.com/blogs/telecom-cybersecurity-is-changing-fast-what-operators-need-to-know-2">Otel 2.0</a> by AT&T.

Our internal asset includes the extent PDF and doc sources for multimodal training, as well as additional data collections for standard bodies like 3gpp or IEEE.

The current sample includes 734 documents (49,954 pages and 15.6 million words, spanning 1999 to 2026) with the original PDF embedded in a single parquet file. Coverage includes 3GPP specifications and working-group contributions, IEEE Access papers, telecom scientific materials NTIA technical reports and World Bank ICT reports.

## Finance

In association with professional organizations in the finance sector in Europe, we have started to collect the largest global corpus of financial regulation and supervision, nearly one million documents in original PDF from central banks, market regulators and international standard-setters.

Alongside international bodies (BIS, FSB, IOSCO, IMF, FATF, IFRS) and EU authorities (ECB, EBA, ESMA, EIOPA, ESRB) we hold the national central banks and regulators of Germany, Spain, Italy, the Netherlands, Ireland, Switzerland, Sweden and France). The corpus is currently extending beyond Europe the collection..

The collection is meant to support the creation of synthetic environments for standard conversion, as global payments are currently undergoing a troublesome migration to ISO 20022.

The current sample includes 5,000 documents in PDF, spanning every region, with associated metadata in parquet.

# Synthetic data seeds/environment

We provide synthetic datasets and environments grounded in open or licenced proprietary seed material, the specifications and the pipelines needed to generate at scale. Each environment is anchored in a real structured asset such demographic distributions, knowledge graphs, etc. 

Our <a href="https://huggingface.co/datasets/PleIAs/data_samples/blob/main/SYNTH__neurips_2026.pdf">synthetic pretraining paper</a> currently under submission to Neurips provides a wider assement of our generalist synthetic pipelines. <a href="https://huggingface.co/datasets/PleIAs/SYNTH>SYNTH</a> (~80B tokens generated from 58,000 seed articles) trains models that sit on the token-efficiency frontier with no SFT or RL stage at all: within 0.7 points of Qwen3-0.6B on open-ended tasks at 80–700× fewer training tokens, and best-in-tier factual precision at 10–140× fewer tokens.

## Global personas.

Our internal assets scaling the synthetic personas at a global scale. We collected a unique corpus of representative first name and last names and aggregated many demographics distribution from international organizations and academic research.

We provide a sample of 100,000 personas across 108 countries. Our full asset includes currently 100 millions personas but is indefinitely scalable. Similarly to Wikidata, we can also provide continuous updates taking into enhanced information and demographic changes.

## Synthetic ethics

We are building the largest structured corpus of framework-conditioned moral reasoning: every trace works through a scenario under one of 16 normative frameworks spanning six ethical traditions, from classical utilitarianism to Kantian ethics, care ethics, moral particularism, pragmatism and constitutional AI principles. Scenarios are generated against a controlled factorial specification (causal structure, harm severity, evitability, number of parties harmed or benefited) and grounded in our global personas asset, so that moral variables can be isolated and balanced the way no scraped corpus allows.

Our internal asset currently includes 54,000 full reasoning traces over 3,400 synthetic dilemma. This current sample features 1,600 traces: 100 synthetic dilemmas each reasoned through all 16 frameworks, with complete factorial metadata, persona grounding and per-trace moral-feature tags.

## Synthetic social media

We built a reusable synthetic pipeline for realistic social media simulations, with a specific focus on expression naturalness/intended noise that was <a href="https://arxiv.org/pdf/2604.18226">presented</a> this year at ACL.

We include a sample of 1,000 synthetic tweets drawing from a large dataset of 2 million tweets developed for one of the leading European public transport operator, RATP. The sample shows the entire process of tweet drafting from our internal specialized models for social media synthetization, including the initial conditioning of nineteen potential features (drawn from a PGM) and an intermediary reasoning phase for staging tweet synthesis. While primarily French, the pipeline is multilingual and support realistic social media expressions in English, German, Hindi, etc.

## Wikidata

Seed-ready latest Wikidata dump at a statements/qualifier level used for the development of a <a href="https://meta.wikimedia.org/wiki/Wikimedia_CH/Grant_apply/AI-BRIDGES:_Synthetic_Reasoning_Corpus">synthetic search environment</a> over Sparql with Wikimedia Deutschland/Wikidata. In contrast with the original dumps, our collection reconcile all ids to labels and restructure the nested statements into a flat structure easily retrievable as parquet. 

The sample is the first file in our collection of 2,369 parquet aggregations. Full dataset includes 600 millions statements and has been used extensively for synthetic environment grounding, diversification and search environment exercises (knowledge traversal, reconciliation). Through our partnership with Wikimedia Foundation, we maintain a regular update.