🚀 I'm a fan of RAG, but there is potentially untapped alpha in fine-tuning, especially in Life Sciences where proprietary data and nuanced terminology are common, the effectiveness of general LLMs can be limited.
With Snowflake Cortex-finetuning, it took me less than 5 minutes to fine-tune the LLaMA-8B, creating a new model, Device_LLaMA_8B, to accurately answer questions based on proprietary sample data in a table.
🔍 The challenge? Analyze raw customer complaint texts from the devices sold by MediSnow (a fictitious company) and determine which Business Unit (BU) manufactured each device.
📉 LLaMA-8B by itself: Small, cheaper to run, but not very powerful.
💪 LLaMA-70B: More powerful, but lacks context. For instance, it has no clue which Business Unit manufactures the QuantumNova X5, a device MediSnow produces and sells.
💡 The solution: Fine-tune the smaller LLaMA-8B, making it both affordable and knowledgeable about MediSnow-specific data.
💬 Again, RAG is a good approach, but remember, you pay for tokens IN and OUT. So, in some cases, front-loading that cost by fine-tuning could be a better approach as opposed to building increasingly complex prompt pipelines with RAG.
🔧 Follow this quick tutorial step-by-step (can be done in under 15 minutes): https://lnkd.in/gcyuXJqQ
Thanks Micha for sharing!