Instructions to use intuitem/Poulpe-Gen1-14B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use intuitem/Poulpe-Gen1-14B with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="intuitem/Poulpe-Gen1-14B")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("intuitem/Poulpe-Gen1-14B")
model = AutoModelForMultimodalLM.from_pretrained("intuitem/Poulpe-Gen1-14B", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use intuitem/Poulpe-Gen1-14B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "intuitem/Poulpe-Gen1-14B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "intuitem/Poulpe-Gen1-14B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/intuitem/Poulpe-Gen1-14B

SGLang

How to use intuitem/Poulpe-Gen1-14B with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "intuitem/Poulpe-Gen1-14B" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "intuitem/Poulpe-Gen1-14B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "intuitem/Poulpe-Gen1-14B" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "intuitem/Poulpe-Gen1-14B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use intuitem/Poulpe-Gen1-14B with Docker Model Runner:
```
docker model run hf.co/intuitem/Poulpe-Gen1-14B
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

Poulpe-Gen1-14B

Poulpe is an open-source language model developed by intuitem and fine-tuned from Mistral's Ministral 3 14B Instruct. It specialises in cybersecurity governance, risk and compliance (GRC), with a focus on the French-speaking community: EBIOS Risk Manager (EBIOS RM), ISO/IEC 27005 and security homologation. It answers in French and English.

Poulpe can serve as the AI backend for CISO Assistant, and works on its own or in any application that speaks the OpenAI chat API. This is Poulpe Gen1, iteration v5; later iterations on the same base keep the Gen1 name and are tagged by iteration.

Intended use

Questions on the EBIOS RM method: workshops, concepts, vocabulary, rating scales, deliverables.
Guiding a risk analysis step by step, from the framing to the risk treatment plan.
Calling tools when an application provides them (for example CISO Assistant's MCP tools), and answering in JSON on request.
General cybersecurity and IT questions.

Out of scope: topics outside cybersecurity and IT (it declines), legal advice, and decisions taken without human review. Answers can be wrong: check them against the ANSSI guide.

Files

Merged weights in safetensors (bf16, about 27 GB), with the chat template and Poulpe's default system prompt (SYSTEM_PROMPT.txt).

Run

from transformers import pipeline

chat = pipeline("text-generation", model="intuitem/Poulpe-Gen1-14B", device_map="auto", torch_dtype="auto")
print(chat([{"role": "user", "content": "Quels sont les cinq ateliers d'EBIOS RM ?"}], max_new_tokens=400)[0]["generated_text"][-1]["content"])

vLLM: vllm serve intuitem/Poulpe-Gen1-14B.

Without a system prompt, the chat template adds Poulpe's own default prompt, the one it was trained with. Recommended temperature: 0.2.

Evaluation

Scores between 0 and 1, held-out sets. Knowledge and facilitation sets are graded by an LLM judge against reference answers and the source pages; tool calling and JSON are checked deterministically. Judge noise is about ±0.02.

Set	Items	This release	v4	v4.1
Closed-book EBIOS RM questions	337	0.405	0.423	0.372
True, false and partial premises	92	0.403	0.407	0.362
Workshop facilitation (multi-turn)	36	0.528	0.517	0.530
Driving a study with terse replies (multi-turn)	8	0.611	0.425	0.595
Identity, scope, general IT	26	0.790	0.762	0.764
Tool calling (CISO Assistant MCP tools, held out)	36	0.986	0.963	0.899
JSON output (schema-validated)	10	1.000	1.000	1.000
JSON with CISO Assistant's system prompt	10	1.000	1.000	1.000

False premises firmly corrected: 12 of 18.
Closed-book answers framed with the wrong method or workshop: 30 of 337.
Repeated replies in the terse-reply sessions: 3 over 8 sessions.
Tool calls claiming an action without calling the tool: 0.

Known limitations

Exact wording of rating scales (gravity, likelihood, stakeholder criteria) and the values of the guide's worked examples can be recalled imperfectly; an example value can be presented as if the method fixed it. Check them against the ANSSI guide, or give the model the relevant pages (retrieval) when precision matters.
Long studies driven with very short replies can still drift or repeat a step.
It follows the EBIOS RM guide v1.5 (2024); the terms of the 2018 version (for example "criticité" of a stakeholder) are not always mapped to the current ones.
Not legal or regulatory advice; a risk analysis remains the responsibility of the people who carry it out.

Training data and sources

Fine-tuned with QLoRA (rank 16) on examples a teacher model (Gemma 4 31B, Apache-2.0) wrote from the documents below: question-answer pairs, corrections of false premises, multi-turn facilitation sessions, plus tool calls and general IT answers produced by the base model. The examples are reformulations, not the documents themselves. Formulas, rating scales and tables that the ANSSI documents only give as figures were transcribed by the teacher from the page images first, and the generated formulas were checked against them. About 40,000 training examples, one epoch on a single RTX 3090.

Contient des informations publiques de l'ANSSI (Agence nationale de la sécurité des systèmes d'information), réutilisées selon la Licence Ouverte / Etalab, avec l'autorisation de l'ANSSI (21 octobre 2026) pour les documents qui ne portent pas de licence.
ENISA reports are reused under CC BY 4.0; © European Union Agency for Cybersecurity (ENISA).
NIST publications are works of the US Government, not subject to copyright in the United States.
MITRE ATT&CK® and CAPEC™ are used under MITRE's licence (see NOTICE-MITRE.txt). "© 2026 The MITRE Corporation. This work is reproduced and distributed with the permission of The MITRE Corporation."

Document	Publisher	Licence
EBIOS RM – Correction de l'étude de cas	ANSSI	Etalab 2.0 (cyber.gouv.fr default; co-produced with Club EBIOS)
Fiche méthode – Ateliers 1 et 3 – Évaluer la gravité des événements redoutés	ANSSI	Etalab 2.0 (confirmed)
Fiche méthode – Atelier 1 – Identifier le périmètre métier et technique	ANSSI	Etalab 2.0 (confirmed)
Fiche méthode – Atelier 2 – Identifier et caractériser les sources de risque	ANSSI	Etalab 2.0 (confirmed)
Fiche méthode – Atelier 3 – Dangerosité des parties prenantes de l'écosystème	ANSSI	Etalab 2.0 (confirmed)
Fiche méthode – Atelier 3 – Définir des mesures de sécurité pour l'écosystème	ANSSI	Etalab 2.0 (confirmed)
Fiche méthode – Atelier 4 – Élaborer des graphes d'attaque	ANSSI	Etalab 2.0 (confirmed)
Fiche méthode – Atelier 4 – Évaluer la vraisemblance des scénarios opérationnels	ANSSI	Etalab 2.0 (confirmed)
Fiche méthode – Atelier 5 – Structurer les mesures de traitement du risque	ANSSI	Etalab 2.0 (confirmed)
Fiche méthode homologation – Identifier l'autorité d'homologation	ANSSI	Etalab 2.0 (pending ANSSI confirmation
Fiche méthode homologation – L'homologation pour les décideurs	ANSSI	Etalab 2.0 (pending ANSSI confirmation
Fiche méthode homologation – L'homologation de sécurité et la norme ISO/IEC 27001	ANSSI	Etalab 2.0 (pending ANSSI confirmation
Fiche méthode homologation – Système d'information isolé et simple	ANSSI	Etalab 2.0 (pending ANSSI confirmation
Fiche méthode homologation – Système d'information répliqué	ANSSI	Etalab 2.0 (pending ANSSI confirmation
Fiche méthode homologation – Organiser une revue des systèmes d'information	ANSSI	Etalab 2.0 (pending ANSSI confirmation
Cartographie du système d'information – Guide d'élaboration en 5 étapes (v1.0	ANSSI	Licence Ouverte / Etalab v1
Le guide de l'homologation de sécurité des systèmes d'information (v2.2	ANSSI	Licence Ouverte / Etalab v2
Guide d'hygiène informatique – Renforcer la sécurité de son système d'information en 42 mesures (v2.0	ANSSI	Licence Ouverte / Etalab v1
La méthode EBIOS Risk Manager – Le Guide (v1.5, 2024)	ANSSI	Licence Ouverte / Etalab v1
Cahier des charges du label EBIOS Risk Manager (v3.1)	ANSSI	Etalab 2.0 (cyber.gouv.fr default)
Panorama de la cybermenace 2025 (mars 2026)	ANSSI	Licence Ouverte / Etalab v2
ReCyF – Référentiel Cyber France v2.5 (document de travail	ANSSI	Etalab 2.0 (ANSSI permission 2026-10-21
EBIOS RM – Support de formation (V2.04	ANSSI	Etalab 2.0 (cyber.gouv.fr default; co-produced with Club EBIOS)
EBIOS RM – Support du stagiaire	ANSSI	Etalab 2.0 (cyber.gouv.fr default; co-produced with Club EBIOS)
Technical implementation guidance on cybersecurity risk management measures (NIS2	ENISA	CC BY 4.0
ENISA Threat Landscape 2025 (v1.3	ENISA	CC BY 4.0
MITRE ATT&CK Enterprise	MITRE	MITRE ATT&CK terms of use (royalty-free
MITRE ATT&CK for ICS	MITRE	MITRE ATT&CK terms of use (royalty-free
MITRE CAPEC (STIX 2.1)	MITRE	CAPEC terms of use (royalty-free
NIST Cybersecurity Framework 2.0 (CSWP 29	NIST	US public domain; NIST reprint permission outside the US
NIST SP 800-30 Rev. 1 – Guide for Conducting Risk Assessments (2012)	NIST	US public domain; NIST reprint permission outside the US

Club EBIOS pages (48), used as practitioner guidance under CC BY 4.0 (site default) or the Creative Commons licence stated in the file

Feedback

Report wrong answers or problems in the CISO Assistant issues, with the question, the answer and the expected one.

Licence

Model weights: Apache-2.0, as the base model. Source documents keep their own licences, listed above.

Downloads last month: 55

Safetensors

Model size

14B params

Tensor type

BF16

Model tree for intuitem/Poulpe-Gen1-14B

Base model

mistralai/Ministral-3-14B-Base-2512

Quantized

mistralai/Ministral-3-14B-Instruct-2512

Finetuned

(11)

this model

Quantizations

2 models