How Do I Evaluate Whether the Fine-Tuned Pangu Model Is Normal?

There are many methods for evaluating model performance, including:

• Loss curve: Evaluate the training effects based on the change trend of the loss curve and check whether exceptions such as overfitting or underfitting have occurred during the training.
• Model evaluation: Use the model evaluation function of the platform to evaluate the test set you have uploaded. You can view metrics such as PPL, BLEU, and ROUGE of test set samples and compare them horizontally (foundation models with different specifications trained on the same data) or vertically (multiple model versions trained on different training data) to determine whether problems have occurred during the training.
• Manual evaluation: You can construct an evaluation set based on the target task and manually evaluate models horizontally or vertically.