Table 13 Diagnostic analysis of instruction prompt sensitivity in Qwen2.5-1.5B-Instruct classification.
From: Qwen TextCNN and BERT models for enhanced multilabel news classification in mobile apps
Prompt type | Prompt template | Acc. | M-F1 | \(\mu\)-F1 |
|---|---|---|---|---|
Specific | Please classify this automotive news article into the appropriate category. Return only the category ID | 0.3064 | 0.0008 | 0.3064 |
Direct | Classify the following automotive news article. Return only the category ID | 0.3063 | 0.0008 | 0.3063 |
Conditional | Given this automotive news article, determine its category. Return only the category ID | 0.3025 | 0.0008 | 0.3025 |
Variant | What category does this automotive news article belong to? Return only the category ID | 0.3064 | 0.0008 | 0.3064 |
Analytical | Analyze this automotive news and provide the category ID | 0.2667 | 0.0008 | 0.2667 |
Performance variance (\(\sigma ^2\)) | 0.0002 | 0.0000 | 0.0002 | |
Coefficient of variation | 5.2% | 0.0% | 5.2% | |
