A training dataset is a collection of task-specific examples used to adapt an existing model. For review classification, each example pairs a review with its desired label, such as neg. section 1.5.2
Fine-tuning data represents successful interactions: an input message followed by the output the model should produce. OpenAI chat models encode each interaction as a message history, with one Python dictionary per message. section 9.8
A practical preparation workflow includes:
- Start with labeled examples, such as rows containing review text and sentiment labels. section 9.8
- Convert each row into a messages object containing a user message and the desired assistant response. section 9.8
- Write one interaction per line in JSON-lines format, commonly using the .jsonl suffix. section 9.8
- Use enough examples to make adaptation useful; the demonstrated range is at least 50 and up to a few thousand samples. section 9.8
Review classification example
A review becomes a training interaction when its text is assigned the user role and its known sentiment, such as neg, is assigned the assistant role. A Python preparation script can read train_reviews.csv and produce train_reviews.jsonl.
