An adaptation technique is the method used to tailor a language model to a task. Options form a continuum from prompting and retrieval-augmented generation (RAG) through LoRA and full supervised fine-tuning (SFT) to distillation and alignment. section 1.11
Choose a technique by matching the task to the method, rather than assuming heavier adaptation is better.
- Ask whether fine-tuning is needed, how much data is available, whether to use an open-source model or frontier API, and whether the goal is fine-tuning, distillation, or alignment. section 1.11
- Treat technique choice as one stage in a lifecycle: base model and curated data, adaptation, evaluation, deployment and monitoring, then feedback to data curation. section 1.11
- Do not expect adaptation alone to fix confabulation; desired behavior must be demonstrated in training data, and preference optimization may also be needed. section 1.11
A costly mismatch in financial analysis
A bank fine-tuned an 8B-parameter model on three years of earnings transcripts, spending eight weeks and $55K in compute. It learned financial language but could not answer questions about the latest call. Because the task required reasoning over live documents, RAG fit better than supervised fine-tuning. section 2.0
