Language Modeling Accuracy on LLM Evaluation Benchmarks (Zero-shot)
68.8Llama-2 7B AccuracyFP16
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| FP16Weight bit=16, Pruning percentage=0%, Reduction size=0×2025.12 | 68.8 | 74 | 72.4 | 74.6 | 76 | 70.6 | |
| UniQLOne pass=true, Fine-tuning (+FT)=true, Weight bit=4, Pruning percentage=0%, Reduction size=4×2025.12 | 67.6 | 73.6 | 72.4 | 75.1 | 73.3 | 69.3 | |
| MoDeGPTOne pass=false, Fine-tuning (+FT)=false, Weight bit=16, Pruning percentage=15%, Reduction size=0.15×2025.12 | 66.2 | 72.4 | 52.1 | — | — | — | |
| UniQLOne pass=true, Fine-tuning (+FT)=true, Weight bit=4, Pruning percentage=15%, Reduction size=4.7×2025.12 | 65.6 | 71.4 | 68.1 | 70.3 | 70.5 | 65.8 | |
| UniQLOne pass=true, Fine-tuning (+FT)=true, Weight bit=4, Pruning percentage=25%, Reduction size=5.3×2025.12 | 63.5 | 67.7 | 64 | 67.4 | 64.7 | 61.8 | |
| SVD-LLMOne pass=false, Fine-tuning (+FT)=true, Weight bit=4, Pruning percentage=15%, Reduction size=4.7×2025.12 | 63.2 | 60.6 | 66.8 | — | — | — | |
| UniQLOne pass=true, Fine-tuning (+FT)=true, Weight bit=3, Pruning percentage=0%, Reduction size=5.3×2025.12 | 62.8 | 64.5 | 67.4 | 67.8 | 71.3 | 67.8 | |
| UniQLOne pass=true, Fine-tuning (+FT)=true, Weight bit=3, Pruning percentage=15%, Reduction size=6.2×2025.12 | 60.2 | 63.5 | 63 | 64.4 | 67.8 | 64 | |
| UniQLOne pass=true, Fine-tuning (+FT)=true, Weight bit=3, Pruning percentage=25%, Reduction size=7.1×2025.12 | 58.9 | 59.4 | 58.7 | 61.6 | 62.1 | 60.3 |