コラム

Many AI models using LLMs, such as ChatGPT, Gemini, and Claude, have emerged recently. These AI models are trained on open data and are designed to be general-purpose.

However, when using them in a business context, it is essential to have knowledge tailored to your company or industry domain. Here, we summarize methods to optimize these LLMs for your own domain.

Fine-Tuning

Fine-tuning is a technique where a pre-trained model is retrained with new data. By adapting the model's parameters to the new data, fine-tuning can enhance the model's performance.

The steps for fine-tuning are as follows:

1. Prepare a pre-trained model
2. Prepare a new dataset
3. Retrain the pre-trained model with the new dataset

The advantage of fine-tuning is that it can achieve high performance on the new dataset. Additionally, fine-tuning is effective even when the new dataset is small.

On the downside, fine-tuning requires as much fresh data as possible. If the new dataset differs significantly from the original dataset, performance may decline. Moreover, fine-tuning requires considerable computational resources, which can be costly.

RAG and Embedding

RAG (Retrieval Augmented Generation) is a model that combines retrieval and generation. It functions like a search engine to retrieve information and then generates text based on that information.

The steps for RAG are as follows:

1. Use a search engine to retrieve information
2. Generate a text (prompt) based on the search results
3. Input the generated text into the AI model to produce results

The advantage of RAG is its ability to use a search engine for information retrieval, leading to more accurate text generation. It also adapts to new information simply by enhancing the search engine's index, eliminating the need for additional training.

The disadvantage is that if the search engine's index contains outdated information, the generated text may also be outdated. Additionally, if the index is inadequate, the quality of the generated text may suffer. Too much input information from search results can also negatively impact the AI model's output, making it highly dependent on the quality of the search results.

Transfer Learning

Transfer learning applies a model trained on one task to another task. By applying knowledge from the original task to a new task, transfer learning can improve the performance of the new task.

The steps for transfer learning are as follows:

1. Prepare a model trained on the original task
2. Prepare a new dataset for the new task
3. Retrain the model on the new dataset

Transfer learning is advantageous when the training data for the business domain is limited and the LLM model is large-scale.

The downside of transfer learning is that if the original and new tasks are significantly different, performance may decline. It also requires significant computational resources, leading to high costs.

RAFT

RAFT is a new method to improve RAG. It trains the model to ignore irrelevant information from the RAG-provided data and select the necessary texts (oracle documents) for generating answers.

It has been shown to deliver better results than traditional fine-tuning and RAG approaches.

via RAFT (Retrieval Augmented Fine-tuning): A new way to teach LLMs (Large Language Models) to be better at RAG (Retrieval Augmented Generation)

CoA

Like RAFT, CoA is a method to improve RAG. CoA enhances performance for questions requiring the combination of multiple pieces of knowledge.

In CoA, the AI splits a question into multiple parts, gathers information externally for each, and then integrates the information to generate an answer.

The advantage of CoA is its high accuracy in answering more complex and diverse questions. By combining multiple pieces of knowledge, CoA can generate more accurate answers.

via [2401.17464] Efficient Tool Use with Chain-of-Abstraction Reasoning

Conclusion

There are several methods to optimize AI models for your company or specific domain. Utilize these techniques to make the most of AI models.

While LLMs offer many possibilities, optimization is essential to make them practical. Be sure to consider other techniques as well.