215 lines
6.7 KiB
Markdown
215 lines
6.7 KiB
Markdown
---
|
||
language:
|
||
- fr
|
||
- en
|
||
license: apache-2.0
|
||
base_model: Qwen/Qwen2.5-0.5B-Instruct
|
||
tags:
|
||
- iso42001
|
||
- ai-governance
|
||
- ai-management-system
|
||
- aims
|
||
- eu-ai-act
|
||
- grc
|
||
- compliance
|
||
- edge
|
||
- qlora
|
||
- mlx
|
||
- fine-tuned
|
||
- on-premise
|
||
- sovereign-ai
|
||
pipeline_tag: text-generation
|
||
library_name: transformers
|
||
---
|
||
|
||
# ISO42001-Qwen2.5-0.5B-Edge
|
||
|
||
<p align="center">
|
||
<img src="https://img.shields.io/badge/ISO%2FIEC%2042001%3A2023-AI%20Management%20System-blue?style=flat-square"/>
|
||
<img src="https://img.shields.io/badge/Base-Qwen2.5--0.5B--Instruct-orange?style=flat-square"/>
|
||
<img src="https://img.shields.io/badge/Fine--tuning-MLX%20%7C%20QLoRA-green?style=flat-square"/>
|
||
<img src="https://img.shields.io/badge/Deploy-On--premise%20%7C%20Offline-purple?style=flat-square"/>
|
||
<img src="https://img.shields.io/badge/License-Apache%202.0-lightgrey?style=flat-square"/>
|
||
</p>
|
||
|
||
> **Specialized SLM for ISO/IEC 42001:2023 — AI Management System.**
|
||
> Fine-tuned on Qwen2.5-0.5B-Instruct. Runs fully on-premise, offline, with no external dependencies.
|
||
|
||
---
|
||
|
||
## Overview
|
||
|
||
| | |
|
||
|---|---|
|
||
| **Base model** | `Qwen/Qwen2.5-0.5B-Instruct` (Apache 2.0) |
|
||
| **Architecture** | Qwen2 — 24 layers · 896 hidden dim · 14 heads · 0.5B parameters |
|
||
| **Fine-tuning** | MLX LoRA (Apple Silicon) · QLoRA 4-bit NF4 (GPU) |
|
||
| **Domain** | ISO/IEC 42001:2023 · EU AI Act · GDPR × AI · AI Governance |
|
||
| **Languages** | French · English |
|
||
| **Deployment** | On-premise · Offline · Ollama · llama.cpp · LM Studio |
|
||
|
||
---
|
||
|
||
## What is this model for?
|
||
|
||
ISO/IEC 42001:2023 is the first international standard for **AI Management Systems (AIMS)**. It provides organizations that develop, deploy, or use AI with a governance framework to demonstrate responsible and ethical AI use — increasingly required in the context of the EU AI Act.
|
||
|
||
This model gives CISOs, DPOs, CAIOs, and GRC consultants precise, clause-referenced answers on:
|
||
|
||
- **Clauses 4–10** — context, leadership, planning, support, operations, performance evaluation, improvement
|
||
- **Annex A** — all controls: A.2 policies · A.6 AI system operation · A.7 transparency · A.8 data governance · A.10 supply chain
|
||
- **EU AI Act × ISO 42001 mapping** — 4 risk levels, obligations per category
|
||
- **ISO 27001 × ISO 42001 × GDPR integration** — unified governance approach
|
||
- **Practical topics** — impact assessment, model cards, SoA, AI system register, privacy risk
|
||
|
||
---
|
||
|
||
## Example queries
|
||
|
||
```
|
||
What is the scope of ISO/IEC 42001:2023?
|
||
How is Annex A of ISO 42001 structured?
|
||
How to conduct an AI Impact Assessment per control A.6.1?
|
||
What are the human oversight requirements under ISO 42001 (A.6.2)?
|
||
How does ISO 42001 map to EU AI Act Article 9?
|
||
What data governance controls does ISO 42001 require for AI systems (A.8)?
|
||
Qu'est-ce qu'un Statement of Applicability dans ISO 42001 ?
|
||
Comment certifier un AIMS ISO 42001 ? Quelles sont les étapes ?
|
||
Quelle est la différence entre ISO 27001 et ISO 42001 ?
|
||
Comment créer un registre des systèmes d'IA conforme à ISO 42001 ?
|
||
```
|
||
|
||
---
|
||
|
||
## Inference
|
||
|
||
### HuggingFace Transformers
|
||
|
||
```python
|
||
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||
import torch
|
||
|
||
model_id = "sallani/ISO42001-Qwen2.5-0.5B-Edge"
|
||
|
||
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
||
model = AutoModelForCausalLM.from_pretrained(
|
||
model_id,
|
||
torch_dtype=torch.bfloat16,
|
||
device_map="auto",
|
||
)
|
||
|
||
messages = [
|
||
{
|
||
"role": "system",
|
||
"content": (
|
||
"You are an expert assistant in AI governance and management systems, "
|
||
"specializing in ISO/IEC 42001:2023 (AI Management System), the EU AI Act, "
|
||
"and GDPR applied to AI. Your answers are precise, clause-referenced, "
|
||
"and tailored to compliance professionals (CISOs, DPOs, CAIOs, GRC consultants)."
|
||
)
|
||
},
|
||
{
|
||
"role": "user",
|
||
"content": "What are the key controls in ISO 42001 Annex A for AI system operations?"
|
||
}
|
||
]
|
||
|
||
text = tokenizer.apply_chat_template(
|
||
messages, tokenize=False, add_generation_prompt=True
|
||
)
|
||
inputs = tokenizer(text, return_tensors="pt").to(model.device)
|
||
|
||
with torch.no_grad():
|
||
outputs = model.generate(
|
||
**inputs,
|
||
max_new_tokens=512,
|
||
temperature=0.1,
|
||
top_p=0.9,
|
||
repetition_penalty=1.1,
|
||
do_sample=True,
|
||
)
|
||
|
||
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))
|
||
```
|
||
|
||
### Ollama (GGUF Q4_K_M)
|
||
|
||
```bash
|
||
ollama create iso42001-edge -f Modelfile
|
||
ollama run iso42001-edge "How to conduct an AI Impact Assessment per ISO 42001 A.6.1?"
|
||
```
|
||
|
||
### llama.cpp
|
||
|
||
```bash
|
||
./llama-cli \
|
||
-m iso42001-qwen2.5-0.5b-q4_k_m.gguf \
|
||
--system-prompt "You are an ISO/IEC 42001:2023 AI governance expert." \
|
||
-p "What is the scope of ISO 42001?" \
|
||
-n 512 --temp 0.1
|
||
```
|
||
|
||
---
|
||
|
||
## Training details
|
||
|
||
### Dataset
|
||
|
||
47 instruction-following Q&A pairs (FR/EN) covering the full standard:
|
||
|
||
| File | Examples | Split |
|
||
|------|----------|-------|
|
||
| `iso42001_train.jsonl` | 37 | Training |
|
||
| `iso42001_test.jsonl` | 10 | Evaluation (out-of-distribution) |
|
||
|
||
**Thematic coverage:**
|
||
|
||
- Clauses 4–6: Context · Leadership · Planning · AI Impact Assessment · Risk assessment
|
||
- Clauses 7–8: Support · Operations · AI lifecycle · Data governance (A.8) · Human oversight (A.6.2)
|
||
- Clauses 9–10: Performance evaluation · Internal audit · Continual improvement
|
||
- EU AI Act × ISO 42001: full 4-level risk mapping
|
||
- ISO 27001 × ISO 42001 × GDPR integration
|
||
- Practical topics: SoA · AI system register · model card · certification steps
|
||
|
||
### Hyperparameters
|
||
|
||
| Parameter | Value |
|
||
|-----------|-------|
|
||
| Technique | MLX LoRA (Apple M-series) |
|
||
| LoRA rank | 8 |
|
||
| LoRA layers | 4 |
|
||
| Iterations | 100 |
|
||
| Batch size | 8 |
|
||
| Learning rate | 5e-5 |
|
||
| Max seq length | 1024 |
|
||
| Optimizer | Adam |
|
||
|
||
---
|
||
|
||
## Offline deployment
|
||
|
||
This model is designed to run **fully locally** with no network calls at inference time.
|
||
|
||
- ✅ No data sent to external cloud services
|
||
- ✅ CPU-compatible via GGUF Q4_K_M (8 GB RAM minimum)
|
||
- ✅ Apple Silicon optimized via MLX
|
||
- ✅ Compatible with Ollama · llama.cpp · LM Studio · Jan
|
||
- ✅ Apache 2.0 license — commercial use permitted
|
||
- ✅ Fully reproducible fine-tuning from source
|
||
|
||
---
|
||
|
||
## Limitations
|
||
|
||
- Compact dataset (47 pairs) — suited for specialized Q&A and evaluation, not production-critical use without further enrichment
|
||
- 0.5B model — limited on complex multi-step reasoning chains
|
||
- Does not replace a certified ISO 42001 audit conducted by a qualified professional
|
||
- Outputs should be reviewed by a subject matter expert before any regulatory decision
|
||
|
||
---
|
||
|
||
## License
|
||
|
||
This model is released under **Apache 2.0**.
|
||
Base model: [Qwen2.5-0.5B-Instruct](https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct) — Apache 2.0, Alibaba Cloud.
|