In a blog post published on August 7, Meta AI's research team outlined key scenarios where fine-tuning can be beneficial, and compared it with other techniques, such as in-context learning and retrieval-augmented generation (RAG).
"With the advent of larger models, the question of fine-tuning has become more nuanced," the researchers noted, highlighting the increased resource requirements and potential pitfalls like catastrophic forgetting for models with over 1 billion parameters.
The guide identifies five archetypes where fine-tuning can be particularly advantageous: tone, style, and format customisation; increasing accuracy and handling edge cases; addressing underrepresented domains; cost reduction through model distillation; and developing new tasks or abilities.
Meta AI's researchers emphasise the importance of considering alternative techniques before resorting to fine-tuning. They recommend experimenting with in-context learning first, due to its simplicity and potential to improve system performance without the need for extensive retraining.
The blog post also provides a nuanced comparison between fine-tuning and RAG, challenging the common notion that RAG should always be tried before fine-tuning. "We think this paradigm is too simplistic," the team states, suggesting that in many cases, a hybrid solution combining both approaches may yield the best results.
To help developers decide between fine-tuning and RAG, the guide offers a series of questions to consider, including the need for external knowledge, custom behaviour, hallucination tolerance, and data dynamics.
Aditya Jain, an Applied Research Scientist at Meta AI, explained, "In most cases, a hybrid solution of fine-tuning and RAG will yield the best results. The question then becomes the cost, time, and additional independent benefit of doing both."
As LLMs continue to evolve and find applications across diverse domains, Meta AI's guidelines provide a valuable resource for developers and researchers seeking to optimise their AI systems through thoughtful adaptation strategies.