Datasets:

PleIAs
/

data_samples

@@ -18,87 +18,37 @@ configs:
     path: personas/*.parquet
 ---
-# Pleias — Evaluation Samples
-Samples of Pleias's non-public seed data, for technical inspection and schema review.
-Volumes below describe what is available beyond these samples.
-| Config | Sample | Available |
-|---|---|---|
-| `epo_patents` | 5,000 patents / 72,662 sections / 102M tokens, with source PDFs | 1.08M patents processed (2020-2025); 8.25M-document full-text backfile 1978-2025 (~48B words) |
-| `eu_science` | 4,704 papers / 69.5M tokens, with source PDFs | 1.43M documents, ~19B tokens, 204 languages |
-| `wikidata_seed` | 1.45M triples / 42k entities | 558M triples, ~34.5 GB |
-| `personas` | 100k personas / 108 countries | 57.3M personas |
-More configs to follow: financial regulation (24 regulators, 12+ jurisdictions) and 3GPP
-specifications.
-Two configs ship the **source PDFs** alongside the processed text, so extraction quality can be
-judged directly and the pairing can be used for document-understanding training.
-## epo_patents
-European Patent Office publications, processed from PDF with structure preserved. One row per
-document **section**, with document metadata repeated on each row (`document_id` regroups them).
-5,000 patents evenly spread across 2020-2025.
-Composition notes: bibliographic front pages, figure lists, and boilerplate disclaimers are
-removed; paragraphs are reassembled across column and page breaks; margin line-numbering is
-dropped. Citation lists are kept but flagged (`is_document_list`). Bibliographic fields come
-from the document itself (INID codes), so `identifier` is the publication number and
-supplementary fields sit in `metadata` as JSON.
-`date` is the publication year. Some documents are republications after opposition, where the
-original grant year differs — that year is preserved in `metadata`.
-**Source PDFs**: `epo_patents/pdfs/epo_pdfs_<year>.tar` (6 files, 3.83 GB). Members are named by
-publication number — `2023/EP1343809B2.pdf` — so they join to `document_id`, even though the
-files as distributed by the EPO are named by application number.
-`epo_patents/epo_5k_pdf_manifest.parquet` maps `document_id -> tar, member, source_file, bytes`.
-## eu_science
-Scientific literature processed from PDF at scale — open-access papers across many languages,
-one row per document. 4,704 papers drawn evenly across the corpus so the language mix is
-representative rather than clustered.
-`language` is empty for ~22% of rows: those documents come from a source pile whose language
-metadata was blank, and the shards they sit in predate language detection. Everything else
-carries a resolved language name.
-**Source PDFs**: `eu_science/pdfs/science_pdfs_<nn>.tar` (5 files, 5.27 GB). Members are named
-`<openalex_id>.pdf`, joining directly to `identifier`.
-`eu_science/science_5k_pdf_manifest.parquet` maps `identifier -> tar, member, bytes`.
-## wikidata_seed
-Wikidata statements parsed into subject-property-value triples with labels resolved and
-qualifiers attached. `value_type` distinguishes entity references from strings, quantities,
-times, coordinates, and monolingual text.
-Rows are ordered so that every statement about the same entity is contiguous — reading
-sequentially yields whole entities (median 22 statements each) rather than scattered triples.
-This is a **curated seed** rather than a raw dump: database cross-reference properties (external
-identifiers such as Commons category, VIAF, GeoNames, Freebase) are removed, as are
-Wikimedia-internal subjects (categories, templates, disambiguation pages) and auto-generated
-numeric items. Unit references inside quantity values are resolved to labels. Cleaning retains
-61% of source rows.
-## personas
-Synthetic population personas across 108 countries, demographically grounded: name, sex, age,
-place (with coordinates and settlement tier), marital status, education, employment status,
-occupation, religion, monthly income in PPP dollars, and health conditions.
-Location is given as place name with coordinates and settlement tier. This sample is drawn
-across the full corpus and shuffled; per-country counts stay proportional to the underlying
-population weighting. `persona_id` is globally unique (`country`-`id`).
-## Licensing
-Samples are provided for inspection and schema review. Commercial training use requires a
-separate written licence. Source material is permissively licensed with document-level
-provenance; per-row `license` and `source_url` are included where applicable.
-Contact: Pierre-Carl Langlais, Pleias.

     path: personas/*.parquet
 ---
+## Multimodal seeds/Pretraining data.
+### Non-US Patents
+While USPTO has been extensively digitized, available public data leave aside non-US patents. We managed to collect the complete collection of European patents (EPO) in the original pdf format, including tens of millions of technical diagram beyond texts.
+The initial sample is split into a parquet file (result of our internal digitization process)
+### Regulated sectors (Finance, telecom)
+We collaborate with leading professional organizations, GSMA and the Authority of
+## Scientific data
+## Structured data seeds
+### Wikidata
+Seed-ready latest Wikidata dump at a statements/qualifier level. 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.
+### 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 1 millions 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 environment
+### Twitter/X
+Synthetic environment prepared for the subway of Paris comprising 2 millions realistic synthetic tweets in French and other foreign language. The synthetic pipeline was presented this year at ACL and is reproducible for multiple similar social media simulations. We especially designed new generators and evaluations for realistic social media emissions.