Sentiment analysis compares a model’s returned output with the expected output and produces a score indicating their alignment. In fine-tuning evaluation, it helps determine whether parameter changes support the intended context. section 12.2.5
Applied during inference testing, it:
- Uses held-out or newly collected instruction-output data that follows the fine-tuning dataset’s structure. section 12.2.5
- Sends instructions to the model, captures responses, and compares those responses with expected outputs. section 12.2.5
- Leaves model parameters unchanged, unlike fine-tuning, and uses the score to guide possible future tuning. section 12.2.5
Scored evaluation with human review
A high sentiment score suggests that tuning met its goals. A low score signals a need to investigate the data or consider another fine-tuning session, while human validation checks outputs that mathematical scoring may misjudge. section 12.2.5
