Legal Finetuning in collaboration with the Mistral models

Using Mistral Models to Explore the Potential of Legal Fine-Tuning

At Jimini AI, we are proud to have collaborated with Mistral AI on the topic of legal fine-tuning. This experience has allowed us to harness the potential of fine-tuning for specialized legal applications, yielding remarkable results.

The first area of improvement concerns legal analysis. By leveraging the BSARD dataset (Belgian Statutory Article Retrieval Dataset), we successfully increased the accuracy of the model’s legal responses by 20%. Concretely, the AI is now capable of providing structured answers with precise references to relevant legal articles, whereas the previous version often produced vague and insufficiently documented responses.

The second aspect of this innovation focuses on legal translation, as illustrated by the French-German case. Using the Multi EURLEX database, which contains 35,000 European legal documents, we developed a specialized translation system with outstanding performance. Tests conducted by bilingual legal experts speak for themselves: in 94% of cases, the translations produced by the new model were rated as superior or equivalent to the original version.

We are currently working on expanding this research to other use cases, particularly in legal analysis and drafting.

This progress marks a major step forward in the intelligent automation of legal work, promising legal professionals a significant gain in efficiency and accuracy.

‍

For more details on this collaboration, read the full article here.

‍