Spaces:

---

HFforLegal

/

README

Like 1

Running

Hello everyone,

I am Ivan Bertrand NGOMEN BISIL, a lawyer from the Cameroon Bar. I would love to get the community's expert feedback on a new instruction fine-tuning dataset I am building for the OHADA jurisdiction (17 French-speaking African countries under unified business laws).

Currently, sovereign African regional laws are heavily under-represented or hallucinated by mainstream LLMs. To solve this, I am designing a strict instruction-following dataset for the AUDCG (Acte Uniforme relatif au Droit Commercial Général).

I have released a 10-row sample on Hugging Face to test the data engineering architecture before scaling to the full dataset :
👉 Bisilivan/dataset-ohada-droit-commercial-general-echantillon

Technical Approach & Mitigating Hallucinations:

• Format: JSONL format with explicit mapping (instruction, input, output).
• Metadata Tagging: Every entry features category, subcategory, difficulty (Beginner to Expert), and entry_type (Definitions, Q&A, and complex Case Studies) for balanced evaluation and training routing.
• Strict Compliance: The target output is designed to never output a legal opinion or solution without explicitly citing the relevant Article and Paragraph from the AUDCG text.

Questions for the Community:

As legal AI experts, I would highly appreciate your critical view on this data structure:

1. Does the current metadata mapping format look optimal for preventing style/fact hallucinations during SFT (Supervised Fine-Tuning)?
2. Given the high density of complex case studies in the sample, would you recommend specific system prompt architectures or standard prompt templates (like Llama-3 or Mistral) to maximize alignment with this JSONL structure?

Thank you so much for your time, your benevolence, and your feedback!

Cc: @HFforLegal @maastrichtlawtech @TristanQUINTARD