-
Identify Complex Tasks: Before finetuning, it's essential to identify tasks that require heavy reasoning and complex problem-solving abilities.
-
Try Existing LLMs: Start by evaluating the performance of existing LLMs like GPT-3.5 or LLAMA on the task. If the model performs satisfactorily, there is no need for finetuning.
Finetuning LLMs requires significant effort, time, and computational resources. It can also lead to overfitting and worse performance if not done carefully. Before proceeding with finetuning, consider the following approaches:
- Provide the LLM with a small number of examples (usually 1-6) that demonstrate the desired task.
- Gradually increase the number of examples, typically up to 6, until the model's performance is satisfactory. After around 6 examples, most of the time it does not improve further. (depend on task and model)
- Also, need to consider that when increasing the number of examples, the token count and context length also matter.
- Prefix Prompting: Prepend the input with a task-specific prefix or instruction.
- Demonstration Prompting: Provide examples of the desired input-output mapping as part of the prompt.
- Chain-of-Thought Prompting: Encourage the LLM to break down complex tasks into a series of intermediate reasoning steps.
- Combine the LLM's language generation capabilities with a separate retrieval component (e.g., a dense vector index or database).
- The LLM can use the retrieved information from external sources to generate more informed and knowledge-grounded outputs.
- RAG models are particularly useful for open-ended question-answering or tasks that require domain-specific knowledge.
If these techniques do not yield satisfactory results, proceed to the finetuning phase.
Finetuning can create a specialized model tailored for the task at hand. In some cases, finetuning alone may not be sufficient. For example:
- Medical Applications: The language used in medical contexts differs from common usage. To capture the nuances and terminology, it may be better to start with pretraining on domain-specific data.
- Case Study: BloombergGPT: This model was pretrained on 51% financial data and 49% common usage data to better understand financial language and concepts.
By following these steps and considering various techniques, you can make an informed decision on whether to finetune an LLM or explore alternative approaches for your specific task.