Search

10 results on this page

🤖

Fine-Tuning Gemma for Complex Legal Reasoning

A practitioner reports that fine-tuning Gemma 4 26B A4B on 100,000 court decisions failed to outperform a prompted base model for generating legal principles. The discussion explores whether the bottleneck lies in data quality, evaluation design, task complexity, or the fine-tuning workflow.

Reddit r/MachineLearningCommunity16h ago#legal-ai#fine-tuning#evaluation
Page 1