Common-sense Reasoning on PIQA, HellaSwag, ARC-C, ARC-E, and OBQA (test)
84.1PIQA AccuracyLLaMA + LoRA
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| LLaMA + LoRAProtocol=Five-Shot, # Params=30B, Model Size=65.1GB, Instruction-tuning=Alpaca2023.05 | 84.1 | 83.3 | 59.5 | 79.2 | 50.6 | 71.4 | |
| LLaMA + PEQAProtocol=Five-Shot, # Params=30B, Model Size=16.9GB, Quantization=4-bit, Instruction-tuning=Alpaca2023.05 | 82.7 | 80.2 | 56.8 | 75.5 | 47.6 | 68.6 | |
| LLaMAProtocol=Five-Shot, # Params=30B, Model Size=65.1GB2023.05 | 82.5 | 82.2 | 56.2 | 74.9 | 47 | 68.6 | |
| LLaMA + LoRAProtocol=Zero-Shot, # Params=30B, Model Size=65.1GB, Instruction-tuning=Alpaca2023.05 | 81.8 | 80.3 | 48.2 | 61.6 | 42.8 | 62.9 | |
| LLaMA + LoRAProtocol=Five-Shot, # Params=13B, Model Size=26.1GB, Instruction-tuning=Alpaca2023.05 | 81.1 | 78.8 | 53.5 | 72.4 | 47 | 66.6 | |
| LLaMA + PEQAProtocol=Five-Shot, # Params=13B, Model Size=7.0GB, Quantization=4-bit, Instruction-tuning=Alpaca2023.05 | 80.7 | 76 | 50.9 | 71.6 | 48 | 65.5 | |
| LLaMA + PEQAProtocol=Zero-Shot, # Params=30B, Model Size=16.9GB, Quantization=4-bit, Instruction-tuning=Alpaca2023.05 | 80.3 | 78.4 | 49.8 | 63.3 | 42.8 | 62.9 | |
| LLaMAProtocol=Zero-Shot, # Params=30B, Model Size=65.1GB2023.05 | 80.1 | 79.2 | 45.5 | 58.9 | 42 | 61.1 | |
| LLaMAProtocol=Five-Shot, # Params=13B, Model Size=26.1GB2023.05 | 80 | 78.4 | 50.4 | 70.8 | 47.2 | 65.4 | |
| LLaMA + LoRAProtocol=Five-Shot, # Params=7B, Model Size=13.5GB, Instruction-tuning=Alpaca2023.05 | 79.9 | 75.2 | 46.4 | 66.5 | 47.2 | 63 | |
| LLaMA + LoRAProtocol=Zero-Shot, # Params=13B, Model Size=26.1GB, Instruction-tuning=Alpaca2023.05 | 79.6 | 76.7 | 46.3 | 62 | 43.2 | 61.5 | |
| LLaMAProtocol=Five-Shot, # Params=7B, Model Size=13.5GB2023.05 | 79.4 | 75.3 | 45.6 | 65.8 | 44 | 62 | |
| LLaMAProtocol=Zero-Shot, # Params=13B, Model Size=26.1GB2023.05 | 79.1 | 76.2 | 44.5 | 59.9 | 42.2 | 60.4 | |
| LLaMA + PEQAProtocol=Zero-Shot, # Params=13B, Model Size=7.0GB, Quantization=4-bit, Instruction-tuning=Alpaca2023.05 | 78.9 | 74 | 46.4 | 62.5 | 42.8 | 60.9 | |
| LLaMA + PEQAProtocol=Five-Shot, # Params=7B, Model Size=3.8GB, Quantization=4-bit, Instruction-tuning=Alpaca2023.05 | 78.9 | 73.2 | 45.1 | 65.4 | 44 | 61.3 | |
| LLaMA + LoRAProtocol=Zero-Shot, # Params=7B, Model Size=13.5GB, Instruction-tuning=Alpaca2023.05 | 78.6 | 73.3 | 43.7 | 55.8 | 43 | 58.9 | |
| LLaMA + PEQAProtocol=Zero-Shot, # Params=7B, Model Size=3.8GB, Quantization=4-bit, Instruction-tuning=Alpaca2023.05 | 77.9 | 71.4 | 42.4 | 57.2 | 42 | 58.2 | |
| LLaMAProtocol=Zero-Shot, # Params=7B, Model Size=13.5GB2023.05 | 77.3 | 73 | 41.4 | 52.5 | 42.4 | 57.3 |