adriencr81/anssi-bge-m3-embeddings

ANSSI cybersecurity guides — BGE-M3 embeddings

Semantic embeddings of 11 ANSSI (French national cybersecurity agency) guides, produced by the open-source project llm-verification-harness.

Project positioning. Transpose aerospace/defense IVVQ (Integration, Verification, Validation, Qualification) practices to non-deterministic RAG/LLM systems. The project's signature deliverable is a Verification Control Document auto-generated per run (Brique 7). This dataset is an intermediate, machine-verifiable artifact from Brique 2.

Contents

• embeddings.npy — float32 matrix of shape (1231, 1024). Each row = one ANSSI corpus chunk encoded with BAAI/bge-m3. L2-normalized at production time, so matrix @ query returns cosine similarity directly.
• embeddings_index_public.jsonl — 1231 JSON lines, row-aligned with embeddings.npy (matrix row i ↔ JSONL line i). Fields per row: doc_id, page_num, chunk_idx, char_start, char_end.

Chunk source text is not included here — see Corpus governance below.

Contract by properties, not bit-for-bit SHA256

A neural embedding is not bit-for-bit reproducible across machines (BLAS/MKL versions, CPU/GPU float sum order, FP non-associativity). Locking embeddings.npy SHA256 as a blocking regression would be a faux contrat — it would fail on the first machine change. The GitHub repo records the hash for traceability but tests verifiable properties:

• matrix.ndim == 2, matrix.shape[1] == 1024, dtype == float32
• ∀i, |‖matrix[i]‖₂ − 1| < 1e-5 — the L2 contract that lets matrix @ query behave as cosine similarity
• matrix.shape[0] == len(embeddings_index_public.jsonl) — row-alignment
• Ordered subset of chunks.jsonl from the GitHub repo (REQ-CHUNK-04 filter: chunks under 10 tokens excluded from the index but kept in the source chunks.jsonl for audit)

Producer environment (pinned)

embeddings.npy was produced under this exact environment, frozen in the repo's corpus/manifest.yaml under derived_artifacts.embeddings_npy.producer_env:

Component	Version
model	BAAI/bge-m3
revision	5617a9f61b028005a4858fdac845db406aefb181
device	cpu
dtype	float32
dim	1024
normalize_embeddings	true
sentence-transformers	5.6.0
torch	2.13.0
transformers	5.13.0
huggingface_hub	1.23.0
numpy	2.5.1

Corpus governance — why the text is not in the index

The source corpus (11 ANSSI guides) is distributed by ANSSI on cyber.gouv.fr under Licence Ouverte 2.0 (Etalab). The project respects that governance by:

• not republishing the source text here — the HF index carries identifiers only (doc_id, page_num, chunk_idx, char_start, char_end), nothing textual.
• publishing the raw chunks (with text) in the GitHub repo under corpus/chunks.jsonl — closer to source, under git audit.
• keeping source PDFs out of the repo (on-demand download with falsifiable SHA256 verification via download_corpus.py).

A user who wants to reconstitute the chunk text from this index can:

1. Clone the GitHub repo adriencr81/llm-verification-harness
2. Load corpus/chunks.jsonl
3. Join on (doc_id, page_num, chunk_idx) between the two files

The gesture "publish the computable artifacts, respect the source corpus governance" is an IVVQ posture: sensitive data never rides in the public artifacts of a verification chain.

Reproduce retrieve() locally

import numpy as np
import json
from sentence_transformers import SentenceTransformer
from huggingface_hub import hf_hub_download

REPO_ID = "adriencr81/anssi-bge-m3-embeddings"

npy_path = hf_hub_download(REPO_ID, "embeddings.npy", repo_type="dataset")
idx_path = hf_hub_download(REPO_ID, "embeddings_index_public.jsonl", repo_type="dataset")

matrix = np.load(npy_path, allow_pickle=False)
index = [json.loads(line) for line in open(idx_path, encoding="utf-8") if line.strip()]

model = SentenceTransformer(
    "BAAI/bge-m3",
    revision="5617a9f61b028005a4858fdac845db406aefb181",
    device="cpu",
)

q = model.encode(
    ["Quelles recommandations MFA de l'ANSSI ?"],
    normalize_embeddings=True,
)[0].astype(np.float32)

scores = matrix @ q
top4 = np.argsort(-scores)[:4]
for i in top4:
    print(f"{scores[i]:.4f}", index[i])

To also get the chunk text field for each hit, use retrieve.py in the GitHub repo — it joins the same matrix with corpus/chunks.jsonl locally.

Downstream use

Consumed by the RAG pipeline in the same project (ask.py, Brique 3): question → retrieval on this matrix → LLM call → citation-parsed answer. The Brique 7 Verification Control Document will report bench results against these embeddings.

License

• Embeddings and public index (this HF dataset): CC-BY-4.0 — attribution required.
• Source corpus text (not redistributed here): remains under ANSSI's Licence Ouverte 2.0 (Etalab).

Cite

Deleuil, A. (2026). llm-verification-harness — Applying aerospace IVVQ
to RAG/LLM systems. https://github.com/adriencr81/llm-verification-harness

Contact

Adrien Deleuil — huggingface.co/adriencr81