finaleads/french-corpus-llm-sample

French Corpus LLM — Sample 500 (v1.4.0)

FINALEADS LLC builds compliance-ready training datasets for French regulated industries. We turn 2.66 billion tokens of French finance, regulatory, and economic open data 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 v1.4.0 release. It is intended for evaluation: format, structure, content quality, license posture, and provenance documentation. The full corpus (~3 million documents / ~2.66 billion tokens) is available under separate terms.

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`](https://app.snowflake.com/marketplace/listing/GZTWZRWEJ2)

• —Brand site: <https://www.frenchcorpus.com>
• —Provider: FINALEADS LLC
• —Release: fpwc-v1.4.0-2026-05-19
• —Specification: see `DATASET_SPECIFICATION_signed.pdf` (PAdES PKCS#7 self-signed, RSA-4096, SHA-256 fe19174ba619486c69d1c48f0dc7019b940870bc527788bf6539e48dfde7d272)
• —Pipeline content hash: 44eae2a94690d64752e1c4548c2c2c015d097dc783f7c5f4e952a38691477280
• —Contact: support@finaleads.com

What's new in v1.4.0

Eleven new finance/regulatory sources added to the v1.3 base via WAF-bypass nodriver extraction (Cloudflare Turnstile / AWS WAF), all GPU-scored on Mistral-7B-Instruct-v0.3 (V100S) :

Combined with the v1.3 base (DILA, EUR-Lex, ACPR, AMF, BdF, DGTrésor, BOFiP, CNIL, ConstIT, ...), the v1.4 corpus reaches 2,998,388 documents / 10.66 GB plain text / ~2.66 B tokens across 27 institutional sources.

Dataset summary

500 documents in French covering statute, jurisprudence, regulator doctrine, and EU-origin legal text. Each document includes the cleaned plain text plus 33 structured fields:

• —Identification: id (SHA-1 of content), source, upstream_id, url
• —Text: text (UTF-8 plain), extracted_at
• —Language: lang (fraLatn), `langscore`
• —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
• —LLM-judge labels (stage 4 Mistral-7B-Instruct): 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 full v1.4 corpus.

Sample composition

The 500-document sample is a stratified preview from v1.3 kept stable across the v1.3 → v1.4 transition (the v1.4 additions are highly specialised finance/regulatory sources that don't change the linguistic mix of the sample). See the previous v1.3 sample card for the per-source breakdown.

The full v1.4 corpus breakdown is documented in STATS.md and DATASET_SPECIFICATION_signed.pdf.

Provenance and licensing

All source data is published under the French Licence Ouverte 2.0 (etalab-2.0), the EU reuse policy (Decision 2011/833/EU), or industry open access (FBF, AMAFI, AFG, France Assureurs — public communications). License compatibility is documented per source in `LICENSE_per_source.md`.

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.
• —Per-document audit trail: the full v1.4 corpus carries 2,998,545 W3C PROV-O JSON-LD records (one per doc) with content SHA-256, licence terms, and AI Act Article 10 compliance block. Available in the full release.
• —Quality scoring methodology: stage 2 CPU scoring (quality_score, quality_tier) and stage 4 LLM-judge scoring (llm_* fields, Mistral-7B-Instruct on V100S) are documented in DATA_DICTIONARY.md.
• —Deduplication: cross-source MinHash LSH at Jaccard threshold 0.7 with k=5 shingles. Drop rates per source are published in DATASET_SPECIFICATION_signed.pdf.
• —Signed Dataset Specification: the bundled PDF is PAdES PKCS#7 self-signed, RSA-4096, with the SHA-256 hash published on this card and on the brand site.

GDPR pseudonymization disclosure

Per-document pseudonymization of personal names in regulator decisions is applied to ACPR (since v1.2) and CNIL deliberations (added in v1.3). Persons are replaced with bracketed aliases ([P1], [P2], ...) with numbering reset per document. Companies, places, and article references are preserved.

This 500-document sample does not include ACPR or CNIL documents (sampling is stratified across the highest-volume sources). Pseudonymization examples are visible in the full corpus.

For the methodology, see the section "Pseudonymization (GDPR safety layer)" in DATASET_SPECIFICATION_signed.pdf.

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"])

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 and case law
• —Comparative analysis of regulator doctrine across institutions (incl. v1.4 additions ESMA / EBA / ACPR / AMF / DGCCRF / TRACFIN)
• —AI Act / GDPR / MiCA / DORA compliance training data (EUR-Lex recent regulations + CNIL deliberations + ACPR sanctions)
• —Data quality methodology research (stage 2 CPU + Mistral-7B LLM-judge fields exposed)

Limitations

• —500-document subset, not a training corpus. Use this sample for format and quality evaluation; the full corpus is sized for training and refresh.
• —v1.4 finance additions absent from sample — the 11 new finance/regulatory sources added in v1.4 are highly specialised and low-volume (16,341 docs total vs. 2.98M base). They are visible in STATS.md and the full corpus; the sample stays at v1.3 stratification for preview stability.
• —EIOPA dropped from v1.4 — EU rate-limiting on europa.eu prevented stable extraction; planned for re-attempt with residential proxy rotation.
• —AFEP dropped from v1.4 — placeholder site, no actionable content.

What's next (v1.5+ roadmap)

• —EIOPA re-extract with proxy rotation
• —DILA legifrance_jorf delta extract (post-2026-05 incremental updates)
• —Pre-computed embeddings (E5-multilingual-large, 1024-dim, INT8 quantized)
• —Semantic changelog per release
• —Per-document AI Act Article 10 attestation rendered as machine-verifiable JSON-LD signed bundle

Citation

@misc{finaleads_french_corpus_llm_2026_v1_4_sample,
  author       = {FINALEADS LLC},
  title        = {French Corpus LLM --- Sample 500 (v1.4.0)},
  year         = {2026},
  month        = {may},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/finaleads/french-corpus-llm-sample},
  note         = {AI Act Article 10 signed Dataset Specification SHA-256 fe19174ba619486c69d1c48f0dc7019b940870bc527788bf6539e48dfde7d272 / pipeline content hash 44eae2a94690d64752e1c4548c2c2c015d097dc783f7c5f4e952a38691477280}
}

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 (~3M documents / ~2.66B tokens, paid subscription, quarterly refresh), see <https://www.frenchcorpus.com/product/> or contact support@finaleads.com.