finaleads/french-corpus-llm-sample

French Corpus LLM — Sample 500 (v1.3.0)

FINALEADS LLC builds compliance-ready training datasets for French regulated industries. We turn 2.6+ billion tokens of French finance, regulatory, legal, and privacy text into audit-trailed, pseudonymized, AI Act Article 10-documented shares — so foundation model and regtech teams can ship into European enterprises without a data-lineage gap.

This is a public sample of 500 stratified documents drawn from the French Premium Web Corpus v1.3.0 release. It is intended for evaluation: format, structure, content quality, license posture, and provenance documentation. The full corpus (2.93 million documents / ~2.6 billion tokens) is available under separate terms via the channels below.

Available on Snowflake Marketplace

Try the full Limited Trial inside Snowflake — 50,000 stratified documents, 7-day evaluation, queryable in-place via a secure share, no download required, auto-fulfilled across all regions :

→ FINALEADS LLC · French Premium Web Corpus — Finance, Regulation & Economic Edition

• Brand site: https://www.frenchcorpus.com
• Provider: FINALEADS LLC
• Release: fpwc-v1.3.0
• Pipeline SHA-256: 2fc1de058fd85f3f3eedc73c1f8a89b571b6e6d70cfa525b202cb3b622f064db
• Contact: support@finaleads.com

What changed in v1.3 vs v1.2

Source	v1.2.0	v1.3.0	Delta
legifrance_jorf	1,250,131	1,868,522	+49 %
jade (Conseil d'État)	0	513,709	new
kali	219,567	198,390	-10 % (after cross-source dedup)
cass (Cour de cassation)	0	143,737	new
eurlex_fr	143,889	100,433	-30 % (cross-source dedup)
legifrance_legi	121,820	64,906	-47 % (cross-source dedup)
cnil (1979-2025)	0	5,135	new
constit (Conseil constitutionnel)	0	3,192	new
dgtresor	50	19,832	+396×
bdf (Banque de France)	65	1,343	+20×
Total	2,148,446	2,932,268	+37 %

Three new pillars introduced in v1.3 :

• Jurisprudence : 657,446 rulings from CASS (Cour de cassation) and JADE (Conseil d'État) — the canonical French judicial decisions.
• Privacy / GDPR doctrine : 5,135 CNIL deliberations 1979 → 2025, including formal sanctions, advisory opinions, and recommendations.
• Constitutional law : 3,192 Conseil constitutionnel decisions.

Combined with the v1.2 finance/regulatory base, v1.3 covers the full French regulatory ecosystem : statute (LEGI), Official Journal (JORF), tax doctrine (BOFiP), prudential supervision (ACPR), market regulation (AMF), monetary policy (BdF), public economics (DG Trésor, INSEE), GDPR/AI Act enforcement (CNIL), constitutional review, and the entirety of judicial jurisprudence.

Dataset summary

500 documents in French covering statute, jurisprudence, regulator doctrine, EU-origin legal text, and privacy/GDPR enforcement. Each document includes the cleaned plain text plus 34 structured fields :

• Identification : id (SHA-1 of content), source, upstream_id, url
• Text : text (UTF-8 plain), extracted_at
• Language : lang (fra_Latn), lang_score
• Length : n_chars, n_words, n_sents, n_tokens_est
• Quality (stage 2 CPU) : avg_word_len, type_token_ratio, upper_ratio, digit_ratio, punct_ratio, line_density, repetition_4gram, repetition_line, boilerplate_score, lexique3_ratio, kenlm_perplexity, quality_score, quality_tier
• Topical labels (stage 3) : is_finance_topical (bool), is_regulatory_topical (bool), sub_vertical (regtech / risque / fiscalite / macro / corporate / autre)
• LLM-judge labels (stage 4b) : llm_coherence, llm_legal_density, llm_ai_slop, llm_finance_value, llm_toxicity, llm_score_composite, llm_tier, llm_raw

See DATA_DICTIONARY.md for the per-field schema and STATS.md for aggregate counts on the sample.

Source distribution in this sample (v1.3 stratified top-decile)

Source	Docs	Description
legifrance_jorf	190	Journal officiel de la République française (laws, decrees, orders)
eurlex_fr	102	EU regulations and directives in French translation
legifrance_legi	63	Consolidated French codes (Civil, Commercial, Monetary, Tax...)
kali	57	National collective bargaining agreements
jade	55	Conseil d'État administrative court rulings
bofip	12	DGFiP tax doctrine
cass	10	Cour de cassation judicial rulings
amf_v2	8	AMF doctrine and sanctions
dgtresor	3	DG Trésor analytical notes
Total	500

The sample is stratified across the top 10 % of documents by llm_score_composite (the Stage 4b LLM-judge aggregate score). Sources outside this top-decile slice (e.g., cnil, constit, acpr, bdf) are present in the full v1.3 corpus but did not make it into this 500-document evaluation slice.

Provenance and licensing

All source data is published under the French Licence Ouverte 2.0 (etalab-2.0), which is compatible with CC-BY-4.0 and the EU reuse policy. See LICENSE.md for the per-source license declarations.

The sample is shipped under etalab-2.0. You may use it for any purpose including commercial use, provided you attribute FINALEADS LLC and the upstream sources.

EU AI Act Article 10 disclosure

This sample ships with documentation supporting the data governance obligations of Article 10 of the EU AI Act :

• Per-document provenance : every record carries source, upstream_id, and url pointing to the canonical authority publication.
• Quality scoring methodology : stage 2 CPU scoring (quality_score, quality_tier) and stage 4b LLM-judge scoring (llm_* fields) are documented in DATA_DICTIONARY.md.
• Topical labelling : documents are classified as is_finance_topical and is_regulatory_topical (Boolean flags) with the methodology documented in the Dataset Specification.
• Deduplication : cross-source MinHash LSH at Jaccard threshold 0.7 with k=5 shingles, 128 permutations. Drop rates per source : 0.3 % (CASS) → 55 % (CONSTIT), with v1.3-wide 27.4 % drop on 4.04 M input documents.
• Pipeline cryptographic anchor : the v1.3 pipeline content hash is 2fc1de058fd85f3f3eedc73c1f8a89b571b6e6d70cfa525b202cb3b622f064db — SHA-256 over every Python file in the extraction → pseudonymization → deduplication → scoring → packaging pipeline.

GDPR pseudonymization disclosure (v1.3 expanded scope)

The v1.3 release extends pseudonymization beyond ACPR (v1.2 scope) to all sources containing natural-person identifiers : Cour de cassation rulings (CASS), CNIL deliberations, AMF sanctions, and ACPR decisions. The ACPR-pattern regex layer replaces party names with bracketed aliases ([P1], [P2], ...) with numbering reset per document. Companies, places, and article references are preserved.

For the methodology, see the section "ACPR pseudonymization (GDPR safety layer)" in the Dataset Specification.

Loading

from datasets import load_dataset

ds = load_dataset("finaleads/french-corpus-llm-sample", split="sample")
print(ds[0]["text"][:500])
print("source:", ds[0]["source"])
print("quality_tier:", ds[0]["quality_tier"])
print("llm_tier:", ds[0]["llm_tier"])
print("llm_score_composite:", ds[0]["llm_score_composite"])

Or with raw jsonlines :

import json
docs = [json.loads(line) for line in open("sample.jsonl", encoding="utf-8")]
print(f"{len(docs)} docs")

Use cases

• Evaluation of French regulatory retrievers (RAG, dense + sparse hybrid)
• Domain-specific SFT data construction for French legal / finance LLMs
• Citation-bearing question answering benchmarks on French statutes, case law, and constitutional decisions
• Comparative analysis of regulator doctrine across institutions
• GDPR / AI Act compliance research — CNIL doctrine is the most authoritative French open source
• Data quality methodology research (stage 2 + stage 4b LLM-judge fields exposed)

Limitations

• 500-document subset, not a training corpus. Use this sample for format and quality evaluation; the full v1.3 corpus is sized for training and refresh.
• Stratified top-decile by LLM-judge score — small sources (CNIL, CONSTIT, ACPR, BdF, AMF) are under-represented in this 500-doc slice but fully present in the production tier on Snowflake.
• EU regulations through 2022 only in eurlex_fr — post-2022 content (DORA, MiCA, AI Act, CSRD) extraction reserved for v1.4 via Playwright crawl.
• No embeddings shipped in v1.3.0. Pre-computed E5-multilingual embeddings are planned for v1.4.

v1.4 roadmap (planned Q3 2026)

• EUR-Lex post-2022 (DORA, MiCA, AI Act, CSRD) via Playwright headless
• Pre-computed embeddings (E5-multilingual-large, 1024-dim, INT8 quantized)
• Per-document distillation rationale (LLM-judge raw output, audit-only column)
• Monthly delta refresh via DILA Légifrance daily diff
• Sub-vertical packaged shares on Snowflake : finance-only / privacy-only / legal-only

Citation

@misc{finaleads_french_corpus_llm_2026_sample_v1_3,
  author       = {FINALEADS LLC},
  title        = {French Corpus LLM --- Sample 500 (v1.3.0)},
  year         = {2026},
  month        = {may},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/finaleads/french-corpus-llm-sample},
  note         = {Pipeline content SHA-256: 2fc1de058fd85f3f3eedc73c1f8a89b571b6e6d70cfa525b202cb3b622f064db}
}

Provider

FINALEADS LLC — premium training datasets for French regulated industries.

• Brand site : https://www.frenchcorpus.com
• Support contact : support@finaleads.com
• Compliance contact : compliance@finaleads.com

For evaluation, the Limited Trial (50,000 docs, 7-day window, free) is live on the Snowflake Marketplace : https://app.snowflake.com/marketplace/listing/GZTWZRWEJ2

For full-corpus access (~2.93 M documents / ~2.6 B tokens, paid subscription, quarterly refresh), see https://www.frenchcorpus.com/product/ or contact support@finaleads.com.