Feature Space Preservation on C4
100Cosine SimilarityBF16
Evaluation Results
| Method | Links | |
|---|---|---|
| BF16Model Architecture=Llama3, Model Size=1B2026.03 | 100 | |
| BF16Model Architecture=Llama3, Model Size=8B2026.03 | 100 | |
| BF16Model Architecture=Qwen3, Model Size=1.7B2026.03 | 100 | |
| BF16Model Architecture=Qwen3, Model Size=8B2026.03 | 100 | |
| FAAR+2FAModel Architecture=Qwen3, Model Size=8B2026.03 | 98.71 | |
| GPTQModel Architecture=Qwen3, Model Size=8B2026.03 | 98.63 | |
| MR-GPTQModel Architecture=Qwen3, Model Size=8B2026.03 | 98.47 | |
| GPTQ+4/6Model Architecture=Qwen3, Model Size=8B2026.03 | 98.42 | |
| FAAR+2FAModel Architecture=Llama3, Model Size=8B2026.03 | 98.39 | |
| 4/6Model Architecture=Llama3, Model Size=8B2026.03 | 98.28 | |
| FAAR+2FAModel Architecture=Qwen3, Model Size=1.7B2026.03 | 98.21 | |
| Strong BaselineModel Architecture=Llama3, Model Size=8B2026.03 | 98.11 | |
| RTNModel Architecture=Llama3, Model Size=8B2026.03 | 98.03 | |
| MR-GPTQModel Architecture=Qwen3, Model Size=1.7B2026.03 | 98.02 | |
| GPTQModel Architecture=Qwen3, Model Size=1.7B2026.03 | 97.95 | |
| 4/6Model Architecture=Qwen3, Model Size=8B2026.03 | 97.91 | |
| Strong BaselineModel Architecture=Qwen3, Model Size=8B2026.03 | 97.82 | |
| GPTQ+4/6Model Architecture=Llama3, Model Size=8B2026.03 | 97.55 | |
| FAAR+2FAModel Architecture=Llama3, Model Size=1B2026.03 | 97.5 | |
| MR-GPTQModel Architecture=Llama3, Model Size=8B2026.03 | 97.41 | |
| GPTQ+4/6Model Architecture=Qwen3, Model Size=1.7B2026.03 | 97.34 | |
| 4/6Model Architecture=Llama3, Model Size=1B2026.03 | 97.33 | |
| Strong BaselineModel Architecture=Qwen3, Model Size=1.7B2026.03 | 97.21 | |
| GPTQ+4/6Model Architecture=Llama3, Model Size=1B2026.03 | 97.14 | |
| RTNModel Architecture=Qwen3, Model Size=1.7B2026.03 | 97.11 | |
| RTNModel Architecture=Qwen3, Model Size=8B2026.03 | 97.11 | |
| MR-GPTQModel Architecture=Llama3, Model Size=1B2026.03 | 97.02 | |
| GPTQModel Architecture=Llama3, Model Size=8B2026.03 | 96.88 | |
| GPTQModel Architecture=Llama3, Model Size=1B2026.03 | 96.01 | |
| Strong BaselineModel Architecture=Llama3, Model Size=1B2026.03 | 95.87 | |
| 4/6Model Architecture=Qwen3, Model Size=1.7B2026.03 | 95.61 | |
| RTNModel Architecture=Llama3, Model Size=1B2026.03 | 94.27 |