Question Answering on PIQA (64 query samples)
82.8AccuracyVictim Model (GPT-3.5-turbo)
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| Victim Model (GPT-3.5-turbo)Model Type=Victim2024.09 | 82.8 | 82.8 | 82.7 | 82.7 | |
| LoRDBase Model=Llama3-8B, Strategy=Locality Reinforced Distillation2024.09 | 78.5 | 79.5 | 78.5 | 78.3 | |
| MLEBase Model=Llama3-8B, Strategy=Model Extraction (MLE)2024.09 | 76 | 77.1 | 76 | 75.7 | |
| KD (gre-box)Base Model=Llama3-8B, Strategy=Grey-box Knowledge Distillation2024.09 | 75.9 | 76 | 75.9 | 75.9 | |
| Local Model (Llama3-8B)Model Type=Local initial model2024.09 | 62.2 | 63.8 | 62.1 | 60.9 |