Fine-tuning takes an existing language model and trains it further on examples of a particular task. Instead of training a model from scratch, you create a specialized variant of a base model. For review sentiment classification, the examples can pair review text with the desired label, such as `pos` or `neg`. section 9.7 section 1.5.2
Try model selection, tuning parameters, and prompt design first. Fine-tuning takes upfront time and money, and using the resulting model can cost more per token. Its potential rewards are better task accuracy and shorter prompts: a model trained to classify reviews may no longer need classification instructions repeated with every review. section 9.0 section 9.7
For a review-classification workflow:
- Start with reviews and their correct sentiment labels. The example training set contains 100 labeled reviews; more samples may help, but increase training cost. section 9.8
- Convert each row into a successful chat interaction: the review is a user message, and its label is the assistant’s reply. Write one interaction per line in a `.jsonl` file. section 9.8
- After fine-tuning, send review text to the resulting model ID and compare its labels with known answers. Because the task is learned, the prompt can be just the review text. section 9.10
{'messages':[
{'role':'user', 'content':'I am willing to tolerate almost anything
...'},
{'role':'assistant', 'content':'neg'}
]}
This sample shows the input and desired output as a user–assistant message pair. section 9.8
import openai
client = openai.OpenAI()
response = client.chat.completions.create(
model='[Fine-tuned model ID]',
messages=[
{'role':'user', 'content':prompt}
]
)
This call uses the fine-tuned model’s ID to request a classification. section 9.10
Check whether it pays off
In a ten-review test, the fine-tuned model got seven labels right and used 2,085 tokens, versus 2,228 tokens for a generic model tuned to the same accuracy. But its stated input-token price was six times higher. Fewer tokens did not offset that price difference in this case; test quality and calculate total cost before deploying. section 9.10
