GPT-3.5 Turbo Fine-Tuning Is Now Available

Fine-tuning for GPT-3.5 Turbo can reduce the need for lengthy instructions and examples, although its higher usage cost may limit its value as a cost-saving measure.

Potential Uses

Fine-tuning is now available for GPT-3.5 Turbo. Previously, it was available only for GPT-3, so GPT-3 was often used when fine-tuning was required. It is good to see this capability added to GPT-3.5 Turbo.

GPT-3.5 Turbo fine-tuning can be used to generate responses with consistent behavior in a desired direction, a required format, or a particular tone of voice.

  • Consistent responses aligned with a desired direction
  • Responses in a desired format
  • Responses in a desired tone

Shorter Prompts and More Context

Until now, these issues could be addressed to some extent through well-written instructions or one-shot and few-shot prompting. With fine-tuning, however, instructions and exemplars may no longer be necessary.

This can shorten prompts and effectively leave more of the context window available for other content. It may be particularly useful when a system prompt has become too long to use comfortably.

Cost Considerations

I initially thought shorter prompts might also reduce costs. However, the pricing makes that difficult to expect.

For a fine-tuned GPT-3.5 Turbo model, training costs $0.008 per 1K tokens. Usage costs $0.012 per 1K input tokens and $0.016 per 1K output tokens.

By comparison, the base GPT-3.5 Turbo model with a 4K context costs $0.0015 per 1K input tokens and $0.002 per 1K output tokens. The 16K-context version costs $0.003 per 1K input tokens and $0.004 per 1K output tokens.

The output price is four times higher than that of the 4K base model, so fine-tuning does not seem likely to provide savings purely through shorter prompts.

  • Fine-tuned GPT-3.5 Turbo: training $0.008 / 1K tokens; input $0.012 / 1K tokens; output $0.016 / 1K tokens
  • GPT-3.5 Turbo, 4K context: input $0.0015 / 1K tokens; output $0.002 / 1K tokens
  • GPT-3.5 Turbo, 16K context: input $0.003 / 1K tokens; output $0.004 / 1K tokens

GPT-4 Pricing for Reference

  • GPT-4, 8K context: input $0.03 / 1K tokens; output $0.06 / 1K tokens
  • GPT-4, 32K context: input $0.06 / 1K tokens; output $0.12 / 1K tokens

Additional Information

Fine-tuning guide: https://platform.openai.com/docs/guides/fine-tuning

Fine-tuning with function calling and fine-tuning for the GPT-3.5 Turbo 16K model are expected to be provided in “late fall.”