This is an LLM performance optimization technique recently revealed at OpenAI's DevDay.
The LLM optimization process is not linear and requires an experimental approach and iterative evaluation.
Initially, we establish a baseline with Prompt Engineering, and apply RAG and fine-tuning as needed.
Through this process, we achieve efficient model performance and apply appropriate optimization techniques depending on the problem type.
Prompt Engineering
Write clear instructions and break complex tasks into simple subtasks. I hear a lot of people talking about long and detailed prompts these days, but in theory, giving clear and concise instructions repeatedly improves reasoning skills.
During this process, it is important to provide clear and concise instructions to minimize token usage.
When complex logical reasoning is required, give the model enough time to think.
Retrieval-Augmented Generation (RAG)
Extend your work by accessing reference texts or external tools. With the advent of GPTs, the concept of RAG has become more important. In the future, it seems that friends like Google's Makesaker, in addition to OpenAI, will actively adopt this method.
RAG updates its knowledge by providing new information to the model.
This method is also used to instruct the model to use only trusted content.
Fine tuning
It provides consistent instructions to the model and emphasizes specialized knowledge needed to solve the problem. This is probably too well known to need explanation.
Fine-tuning is a technique for additionally training an existing model with a domain-specific dataset.
This allows us to improve the performance and efficiency of our models, while maximizing performance with less data constraints.
Model evaluation?
Recently, we often see cases where small models or forked models are trained separately on evaluation datasets to only increase scores. This is not only ethically wrong, but can also have a negative impact on the evaluation methods of other models.
When evaluating the performance of a model, we consider metrics such as accuracy, stickiness, and response appropriateness.
When using the RAG model, the suitability of the content is also an important consideration.
In the case of OpenAI's GPT-4, the case where 98% accuracy was achieved was improved by various attempts such as re-ranking, rule-based methods, and classification. These cases show that the problem can be solved with appropriate context and selection even without fine-tuning.
Subscribe to 'haebom'
📚 Welcome to Haebom's archives. --- I post articles related to IT 💻, economy 💰, and humanities 🎭. If you are curious about my thoughts, perspectives or interests, please subscribe. Would you like to be notified when new articles are posted? 🔔 Yes, that means subscribe. haebom@kakao.com
