How Do I Determine Whether to Adjust Prompts or Use Scenario-Specific Fine-Tuning?

Consider the following two aspects to determine whether to adjust prompts or use scenario-specific fine-tuning:

• Availability of business data

  Check whether the service data of the task scenario is publicly available. If the data related to the scenario can be publicly obtained, the model may have been exposed to similar corpora during the training phase, and therefore can understand the data to a certain extent. In this case, you can adjust the prompt to guide the model to generate reasonable answers.

  For example, for some common Q&As, such as questions of general knowledge, the model can understand and generate accurate answers because related data in these fields is widely available. In this case, you can adjust the prompt to guide the model to generate the required information in a proper style.

• Service logic complexity

  Determine whether the service logic of the task scenario complies with the general logic. If the service logic in the scenario is simple, common, and easy to understand, you can adjust the prompts.

  For example, for common Q&A scenarios, you can adjust the prompts to guide the model to learn how to answer questions concisely and clearly.

  If the scenario involves complex and professional service logic, such as financial analysis and medical diagnosis, a more precise processing method is required.

  • If the service rules in the scenario are simple and easy to summarize, you can use the few-shot method to provide a few examples to the model, so that the model can understand the task and perform inference.
  • If the service rules are complex and difficult to summarize, you are advised to use the scenario-based fine-tuning method to train the model for the specific scenario, so that the model can better understand and adapt to these complex rules.