Zero-shot Language Model Evaluation on MobileLLM Evaluation Suite (350M)
41.86Average PerformanceEDGERAZOR
Evaluation Results
| Method | Links | |
|---|---|---|
| EDGERAZORW-A-KV=4-8-8, Group Sizes=256, Training Budget=1.2B2026.04 | 41.86 | |
| ParetoQW-A-KV=4-16-16, Group Sizes=channel, Training Budget=10B2026.04 | 40.96 | |
| EfficientQATW-A-KV=4-16-16, Group Sizes=64, Training Budget=33M2026.04 | 40.89 | |
| EDGERAZORW-A-KV=2.79-8-8, Group Sizes=64, Training Budget=2.4B2026.04 | 40.62 | |
| ParetoQW-A-KV=3-16-16, Group Sizes=channel, Training Budget=10B2026.04 | 40.24 | |
| EfficientQATW-A-KV=3-16-16, Group Sizes=64, Training Budget=33M2026.04 | 39.27 | |
| EDGERAZORW-A-KV=1.88-8-8, Group Sizes=64, Training Budget=3.1B2026.04 | 39.02 | |
| ParetoQW-A-KV=2-16-16, Group Sizes=channel, Training Budget=30B2026.04 | 38.99 | |
| EDGERAZORW-A-KV=1.58-8-8, Group Sizes=64, Training Budget=3.1B2026.04 | 38.12 | |
| ParetoQW-A-KV=1.58-16-16, Group Sizes=channel, Training Budget=30B2026.04 | 38 | |
| EfficientQATW-A-KV=2-16-16, Group Sizes=64, Training Budget=33M2026.04 | 36.24 |