[2505.04388] The Aloe Family Recipe for Open and Specialized Healthcare LLMs

0 2 minutes read

250504388 The Aloe Family Recipe for Open and Specialized Healthcare.png

[Submitted on 7 May 2025]

Authors:Dario Garcia Gazola, Jordi Bayari-Palianas, Ashwin Kumar Goragan, Enrique Lopez Quina, Adrian Tormos, Daniel Hingos, Pablo Bernabe Breeze, Anna Arias Duart, Pablo Agostin Martin Torres, Marta Gonzalz Malu Ayguadé-Parra, Ulises Cortés

View the PDF file from the paper entitled Aloe Family Recipe for Open and Specialized Healthcare, by Dario Garcia-Gasulla and 12 other books

PDF HTML (experimental) view

a summary:The purpose: With the progress of the LLMS for health care, the need for open source competitive models to protect the public interest. This work in the open medical LLMS field contributes to improving the main stages of pre -treatment of data and training, while showing how to improve the safety of the model (through DPO) and effectiveness (through RAG). The evaluation methodology used, which includes four different types of tests, determines a new standard for the field. The resulting models, which are found to be competitive with the best special alternatives, are released with a coffee license.

Methods: At the top of strong basic models such as Llama 3.1 and QWEN 2.5, Aloe Beta uses a set of data dedicated to enhancing general data with a series of artificial thought examples. Models are subject to compatibility with improved direct preference, while emphasizing moral and political performance in the presence of protection fracture attacks. The evaluation includes close and open assessments, safety and human assessments, to increase the reliability of the results to the maximum.

Results: Recommendations are made through the entire pipeline, with the support of the solid performance of the aloe family. These models provide competitive performance through health care standards and medical fields, and are often preferred by healthcare professionals. On prejudice and toxicity, aloe vehicle models are greatly improving safety, indicating resilience in invisible attacks. As for the official release, a detailed evaluation of the risk of health care is attached to the aloe family models.

Conclusion: Aloe models, the beta, and the recipe that leads to it, are a great contribution to the open source LLM medical field, providing great performance while maintaining high moral requirements. This work determines a new standard for the development of LLMS reporting in the field of health care.