Why does my fine-tuned GPT-3.5 Turbo model output duplicated and incomplete responses with function calling?

Below are some avenues to consider when troubleshooting the integration of a custom persona with function calling:

1. Separation of Persona and Function Structure

When you add function calling, the model is now balancing two roles: following your persona instructions and formatting responses for function calls. The structured output required by function calls might override or conflict with the persona-specific language. Consider whether you really need the function output to include the persona directly—or if you can have a separate step where the persona is applied after the function call response is generated.

2. Clear Function Descriptions

Make sure that the function definitions are as descriptive as possible. They should indicate when personality-specific language is expected. If your fine-tuning examples don’t clearly show how the persona should interact with function output, the model may default to following strict formatting rules meant for function calling rather than maintaining its character.

3. Adjusting Fine-Tuning Examples

Double-check the distribution of examples in your fine-tuning dataset. If the persona was working well in one set but then degrades when adding function calling examples, you may need to:

• Include a greater proportion of examples where the persona is maintained even within function call outputs.
• Ensure that the function description and the response examples contain explicit mention of persona traits or instructions on how to weave the persona into the response.

4. Model Prompting and Context

Consider whether the prompt provided at runtime (and any system instructions) mentions the persona consistently. When introducing function calls, ensure the conversation history still contains references to the persona. This might require adjusting your runtime context or synthesizing a “pre-message” that reminds the model of its character before processing the function call.

5. Experiment with Temperature and Sampling Parameters

Sometimes the issue of duplicated responses or forgotten context can be mitigated by fine-tuning sampling parameters. For example, a lower temperature or adjusted top_p might help the model better adhere to constraints. Test a few settings to see if the output quality improves while keeping both the persona and the function structure.

Final Thoughts

Mixing specialized persona outputs with structured function calling can be tricky, as the model must balance adherence to formatting with the previously trained style. A potential approach is to decouple these responsibilities if possible—first performing the function call, then applying persona-specific reformatting if needed. This two-step process might yield more consistent results.

Keep iterating on your fine-tuning examples with explicit instructions on both fronts to achieve the desired balance.