Model Details
Model Description
Sexo-FR is an instruction-tuned causal language model specialised for the sexual-health domain in French. It was fine-tuned from a general-purpose open base model on a curated corpus of French-language question/answer and instruction data covering topics such as contraception, sexually transmitted infections (IST), consent, sexual and reproductive health, relationships, sexual well-being, and inclusive information on gender identity and sexual orientation.
The training emphasised three things: (1) factual reliability, with content aligned to recognised public-health and clinical guidance; (2) a caring, non-judgemental and inclusive tone; and (3) safe behaviour, including refusing to produce sexually explicit content, recognising distress or emergencies, and systematically signposting appropriate professional resources and helplines.
- Developed by: [French public-health body — e.g., Santé publique France — to confirm]
- Funded by [optional]: Government of France — [e.g., Ministère de la Santé et de la Prévention — to confirm]
- Shared by [optional]: [Same as above / project team — to confirm]
- Model type: Decoder-only (causal) transformer language model, instruction-tuned for domain-specific conversational use
- Language(s) (NLP): French (
fr)
- License:
apache-2.0 (inherited from the base model). Note: documentation, prompts, and any released data may additionally be published under the French government open licence (Licence Ouverte / Etalab 2.0). Confirm the final licensing strategy with your legal team before release.
- Finetuned from model [optional]: [
mistralai/Mistral-7B-Instruct-v0.3 — confirm exact base model and revision]
Model Sources [optional]
Uses
Direct Use
Sexo-FR is intended to power a French-language assistant that answers general sexual-health questions and points people to appropriate care. Representative in-scope topics include:
- Contraception: available methods, emergency contraception, where and how to obtain them.
- Sexually transmitted infections (IST): prevention, awareness of symptoms, screening (dépistage), prevention tools (condoms, PrEP), and how to seek testing or treatment.
- Consent and healthy relationships (consentement): understanding consent, communication, recognising coercion or violence.
- Sexual and reproductive health: anatomy and physiology, pregnancy, contraception, and information about IVG (voluntary termination of pregnancy).
- Sexual well-being and common difficulties: general, non-clinical information about desire, pleasure, pain, or function — always with a recommendation to consult a professional for personal situations.
- Life stages: puberty and adolescence, menopause, and sexuality with age, disability, or chronic illness.
- Identity and inclusion: respectful, inclusive information regarding sexual orientation and gender identity (LGBTQ+).
- Prevention and support: recognising sexual violence and signposting dedicated support services.
Typical deployments: a chatbot or Q&A feature on an official public-health website, an FAQ-assistance layer, or an internal drafting aid reviewed by human experts.
Downstream Use [optional]
The model can be further adapted for specific public-health programmes, for example:
- Fine-tuning on additional, expert-validated content for a particular campaign or audience (e.g., youth, new parents, specific clinical pathways).
- Integration into a retrieval-augmented (RAG) system grounded in an authoritative, regularly updated knowledge base so that answers cite up-to-date official sources.
- Use as an assistant for trained counsellors and helpline operators, with human review before any message is sent to a member of the public.
Any downstream use should preserve the safety behaviours described below and undergo renewed evaluation.
Out-of-Scope Use
This model must not be used for:
- Medical diagnosis, treatment decisions, prescriptions, or dosing. It is not a medical device and provides general information only.
- Emergency or crisis situations. It is not designed to handle medical emergencies, acute distress, suicidal ideation, or ongoing violence/abuse. Such cases must be routed to emergency services and dedicated helplines.
- Replacing professional care by a doctor, midwife, sexologist, psychologist, or family-planning service.
- Generating sexually explicit, pornographic, or arousal-oriented content. The model is for health information and education, not erotica.
- Unsupervised use with minors. Deployments accessible to adolescents must align with national sexual-education standards, remain age-appropriate, and route to youth-appropriate services. The model must never be used to facilitate any sexualisation of, or harm to, minors.
- Legal advice (e.g., precise statutory or regulatory determinations).
- Use outside the French language and the French health/care context without separate validation.
- Profiling, surveillance, or any re-identification of users based on the sensitive content of their questions.
Bias, Risks, and Limitations
This model operates in a sensitive and high-stakes domain. Incorrect, incomplete, or poorly framed information about sexual health can cause real harm. The following limitations are inherent and must be communicated to users and deployers.
Technical limitations
- Hallucination and factual error. Like all LLMs, the model can produce plausible-sounding but incorrect statements, including about medical facts. It can also reflect outdated guidance, since its knowledge is frozen at training time and recommendations evolve.
- No source guarantees. Unless deployed with retrieval grounding, the model cannot reliably cite or verify sources.
- Sensitivity to phrasing. Answers may vary with prompt wording; ambiguous questions may be misinterpreted.
- Language scope. The model targets standard French; it may handle regional variants, slang, mixed languages, or non-French input less reliably.
Sociotechnical risks and biases
- Cultural and demographic bias. Training data may over-represent certain perspectives and under-represent others (e.g., LGBTQ+ experiences, people with disabilities, diverse cultural or religious contexts). The model may carry heteronormative or other normative biases.
- France-centric framing. Guidance, services, and helplines referenced may be specific to France and not transferable elsewhere.
- Tone and stigma. Despite tuning for a non-judgemental tone, the model could occasionally produce content that feels stigmatising or insufficiently inclusive.
- Privacy and sensitivity. Users may disclose highly intimate or identifying information. Deployers are responsible for data minimisation, secure handling, and compliance with the GDPR and applicable French regulations.
- Over-reliance. Users may treat answers as personal medical advice. Clear, persistent disclaimers and easy routes to human help are essential.
Recommendations
Users (both direct and downstream) should be made aware of the risks, biases, and limitations of the model. In addition:
- Always present the model as informational, not a substitute for professional care, with a visible disclaimer at the point of use.
- Keep a human in the loop for any individual situation, clinical question, or message sent to a vulnerable user.
- Prefer retrieval-grounded deployment against an authoritative, current knowledge base, and review content periodically against the latest official guidance.
- Surface professional resources and helplines prominently, and verify these contacts are current for your jurisdiction. Examples of French resources to consider signposting (please verify numbers before release):
- Emergencies: 15 (SAMU), 112 (European emergency number), 114 (emergency by SMS, for deaf/hard-of-hearing).
- Suicide prevention: 3114 (national prevention line).
- Violence against women: 3919 (Violences Femmes Info).
- Child protection: 119 (Allô Enfance en Danger).
- Youth sexual-health information: Fil Santé Jeunes.
- General sexual-health information: a public, official sexual-health information service (e.g., a
questionsexualite.fr-type resource).
How to Get Started with the Model
The model uses a standard chat/instruction format. A system prompt should encode the informational-only, caring, and safety-first behaviour. Always verify the chat template of your chosen base model.
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "your-org/sexo-fr"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype=torch.bfloat16,
device_map="auto",
)
system_prompt = (
"Tu es un assistant d'information en santé sexuelle conçu dans le cadre "
"d'une initiative de santé publique française. Tu fournis des informations "
"fiables, bienveillantes, inclusives et fondées sur des données probantes, "
"en français clair. Tu ne poses jamais de diagnostic et tu ne remplaces pas "
"un professionnel de santé. Tu ne produis aucun contenu sexuellement "
"explicite. En cas d'urgence, de détresse ou de violence, tu orientes vers "
"les services et numéros d'aide appropriés."
)
messages = [
{"role": "system", "content": system_prompt},
{"role": "user", "content": "Quelles sont les principales méthodes de contraception ?"},
]
inputs = tokenizer.apply_chat_template(
messages, add_generation_prompt=True, return_tensors="pt"
).to(model.device)
outputs = model.generate(
inputs,
max_new_tokens=512,
temperature=0.3,
top_p=0.9,
do_sample=True,
)
print(tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True))
A low temperature (e.g., 0.2–0.4) is recommended to favour factual, conservative answers in this domain.
Training Details
Training Procedure
Training Hyperparameters
The following are representative values for a parameter-efficient (LoRA/QLoRA) supervised fine-tune on L40 GPUs — replace with your actual configuration:
- Training regime: bf16 mixed precision
- Method: Supervised fine-tuning (SFT) with LoRA (optionally an additional preference-optimisation pass, e.g., DPO, for safety/tone)
- LoRA config (example): rank
r = 16, alpha = 32, dropout 0.05, applied to attention/MLP projection layers
- Optimizer: AdamW (paged 8-bit if using QLoRA)
- Effective batch size (example):
32 (per-device batch × gradient accumulation)
- Max sequence length (example):
2048
Evaluation
Factors & Metrics
Factors
Evaluation should be disaggregated by, at minimum:
- Topic (contraception, IST, consent, reproductive health, identity, well-being, etc.)
- Question type (factual, procedural/"where do I go", sensitive/personal, safety-critical)
- Population/inclusivity (e.g., LGBTQ+ relevant queries, different life stages)
- Risk category (general information vs. safety-critical vs. out-of-scope)
Metrics
- Factual accuracy / clinical correctness — expert rating of answers against recognised guidance (primary metric in this domain).
- Safety — rate of appropriate refusals for explicit-content/harmful requests and appropriate redirection to professionals/helplines; absence of explicit content.
- Appropriate signposting — proportion of answers that direct users to professional care or relevant services when warranted.
- Helpfulness & clarity — readability and usefulness, rated by reviewers.
- Inclusivity & tone — non-judgemental, inclusive language.
- Coverage — performance breadth across topics.
Summary
[Add a short narrative summary of strengths and known weaknesses once evaluation is complete.]
Model Examination [optional]
[Optional.] Any interpretability or qualitative-analysis work — e.g., review of failure cases, audits of bias on inclusivity-relevant queries, or red-teaming findings — can be summarised here.
Environmental Impact
Carbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).
The figures below are illustrative; replace them with your measured values.
- Hardware Type: NVIDIA L40 (48 GB GDDR6, ~300 W TDP)
- Hours used: [e.g., 192 GPU-hours — to confirm]
- Cloud Provider: [e.g., on-premise / French sovereign-cloud (SecNumCloud) provider — to confirm]
- Compute Region: France
- Carbon Emitted: [e.g., ~4–5 kg CO₂eq — illustrative]
Note on the estimate: France's electricity grid is relatively low-carbon (broadly on the order of ~55–60 gCO₂eq/kWh, varying over time). For example, ~192 GPU-hours at ~0.3 kW, with a data-centre PUE of ~1.3, gives roughly ~75 kWh and on the order of ~4 kg CO₂eq. Substitute your real GPU-hours, PUE, and grid intensity for a precise figure.
Technical Specifications [optional]
Model Architecture and Objective
Decoder-only transformer (causal language model), trained with a standard next-token prediction objective during instruction fine-tuning. [Confirm parameter count and architecture from the chosen base model — e.g., ~7B parameters for Mistral 7B.]
Compute Infrastructure
Hardware
NVIDIA L40 GPUs.
Software
PyTorch, Hugging Face transformers, peft (LoRA/QLoRA), trl for SFT/DPO, accelerate/deepspeed for distributed training.
APA:
[Developing organisation]. ([Year]). Sexo-FR: A French sexual-health information language model [Computer software/model]. Hugging Face. https://huggingface.co/your-org/sexo-fr
Glossary [optional]
- IST — Infections sexuellement transmissibles (sexually transmitted infections, STIs).
- IVG — Interruption volontaire de grossesse (voluntary termination of pregnancy).
- PrEP — Pre-exposure prophylaxis (HIV prevention).
- Dépistage — Screening/testing.
- Consentement — Consent.
- RAG — Retrieval-augmented generation: grounding model answers in an external, authoritative knowledge base.
This model is an informational and educational tool only. It does not provide medical advice, diagnosis, or treatment and does not replace consultation with a qualified professional. In an emergency, contact local emergency services. [Add programme details, governance, and contact channels here.]
[francois@chastel.co]