Table 6.

The error analysis table. All models show a high level of spurious error, but fine-tuning substantially reduces the error rate, showing the lowest total error counts among other methods.

Model versions	Spurious error, n	Granularity error, n	Misalignment error, n	Total error count, n
Vanilla model	76	7	37	120
Fine-tuned	46	1	33	80
Augmented dataset	80	2	33	115