Why a dedicated AML legal dataset?
General-purpose AI systems are neither designed nor trained to reliably meet the specific legal and regulatory requirements of anti-money laundering and counter-terrorist financing (AML/CFT).
AML_LUX_DATASET v2.0.0 addresses a precise need:
- to assess the actual legal compliance of AI-generated responses,
- to measure the risk of out-of-corpus hallucinations,
- to objectively compare multiple models or RAG configurations,
- to document AI governance within an AI Act / internal control framework.
This dataset is not a simple question–answer collection:
it is a legally constrained AI audit tool.
Regulatory scope covered
The dataset is entirely based on a documented Luxembourg and European legal corpus, including in particular:
- Amended Law of 12 November 2004 (AML/CFT)
- Laws governing the Financial Intelligence Unit (FIU) and criminal sanctions
- CSSF circulars (12/02, 17/650, 18/702, etc.)
- European AML Directives (4th, 5th, and 6th AML Directives)
- International standards and recommendations (FATF)
📌 No response is produced outside this defined corpus.
Technical features of the dataset
Grounded and traceable dataset
Each answer is:
- generated under strict corpus constraints,
- supported by explicit citations,
- linked to normalized legal sources,
- structured for machine use (JSONL format).
Hallucination mitigation
The dataset includes:
- insufficient-context cases,
- deliberately blocked responses,
- an explicit and documented refusal logic.
➡️ Ideal for testing whether an AI system knows when not to answer.
Main use cases
🔍 Audit and benchmarking of legal AI systems
- Compare GPT, Claude, Mistral, and internal LLMs
- Test different RAG architectures
- Measure the regulatory robustness of AI outputs
🧠 Training and evaluation of local models
- Controlled fine-tuning
- Post-training evaluation
- Detection of out-of-corpus drift
💬 Compliance & AML chatbots
- Internal compliance assistants
- AML regulatory assistants
- Decision-support tools (non-decision-making)
📊 AI governance & AI Act compliance
- Documentation of AI risks
- Proof of information perimeter control
- Support for internal and external audits
Format & integration
- Format: JSONL
- Language: French legal language
- Version: v2.0.0 (frozen dataset)
- Compatibility:
- RAG frameworks (Chroma, FAISS, Pinecone…)
- Internal AI pipelines
- BULORA.ai audit tools
License & conditions of use
- Internal professional use only
- Redistribution prohibited
- No training of public AI models
- Contractual license provided with the dataset
➡️ See the Offers & Licenses page
Integration with BULORA.ai
AML_LUX_DATASET v2.0.0 is natively compatible with BULORA.ai modules:
- Reasoning
- Source
- Robustness
- Temporal
- Disagreement
It can also be used independently of the platform.
Access & demonstration
Would you like to test this dataset on your own models or use cases?
➡️ Contact us for a demo: contact(@)bulora.ai
➡️ Request an evaluation access: contact(@)bulora.ai
