I really enjoyed reading your story. Your explanation of balancing training efficiency with model safety across multiple epochs was especially interesting. It’s great to see practical approaches that make large-scale embedding training more reliable and efficient. What was the biggest challenge your team faced when implementing this approach?