Commonsense Reasoning on ARC-e, ARC-c, OBQA, PIQA, HellaSwag (Zero-shot)
60.34Average AccuracyLongChat-7B-v1.5
Evaluation Results
| Method | Links | |
|---|---|---|
| LongChat-7B-v1.5Bit=16, Compression Rate=1×, Inference Mode=zero-shot2026.06 | 60.34 | |
| KIVI-2-gs32-r32Bit=3.16, Compression Rate=5.1×, Inference Mode=zero-shot2026.06 | 60.3 | |
| STAR-KV (60%)Bit=3.2, Compression Rate=12.5×, Inference Mode=zero-shot2026.06 | 60.12 | |
| STAR-KV (75%)Bit=3.2, Compression Rate=20.0×, Inference Mode=zero-shot2026.06 | 59.18 | |
| KVQuant-2bit-1%Bit=2.33, Compression Rate=6.9×, Inference Mode=zero-shot2026.06 | 58 | |
| Palu (50%)Bit=4, Compression Rate=8.0×, Inference Mode=zero-shot2026.06 | 56.71 | |
| Palu (50%)Bit=3, Compression Rate=10.4×, Inference Mode=zero-shot2026.06 | 56.42 |