List your company

Commonsense Reasoning on HellaSwag (Robustness Evaluation)

82.94Accuracy (Baseline)BF16
24.4439.627554.81570.0025Dec 14, 2025Dec 23, 2025Jan 2, 2026Jan 12, 2026Jan 22, 2026Feb 1, 2026Feb 11, 2026

Evaluation results

MethodLinks
BF16Model=Qwen2.5-14B, A-W Quant Type=BF162026.0282.94——
NVFP4+PTSModel=Qwen2.5-14B, A-W Quant Type=NVFP4+PTS2026.0281.45—-1.49
HiF4Model=Qwen2.5-14B, A-W Quant Type=HiF42026.0281.45—-1.49
NVFP4Model=Qwen2.5-14B, A-W Quant Type=NVFP42026.0281.38—-1.56
HiF4+HiGPTQModel=Qwen2.5-14B, A-W Quant Type=HiF4+HiGPTQ2026.0281.38—-1.56
BF16Model=Mistral-7B, A-W Quant Type=BF162026.0280.5——
NVFP4+PTSModel=Mistral-7B, A-W Quant Type=NVFP4+PTS2026.0279.71—-0.79
HiF4+HiGPTQModel=Mistral-7B, A-W Quant Type=HiF4+HiGPTQ2026.0279.48—-1.02
HiF4Model=Mistral-7B, A-W Quant Type=HiF42026.0279.41—-1.09
BF16Model=Llama3-8B, A-W Quant Type=BF162026.0279.15——
HiF4Model=Llama3-8B, A-W Quant Type=HiF42026.0277.84—-1.31
NVFP4+PTSModel=Llama3-8B, A-W Quant Type=NVFP4+PTS2026.0277.6—-1.55
HiF4+HiGPTQModel=Llama3-8B, A-W Quant Type=HiF4+HiGPTQ2026.0277.47—-1.68
NVFP4Model=Llama3-8B, A-W Quant Type=NVFP42026.0276.56—-2.59
BF16Model=Llama2-7B, A-W Quant Type=BF162026.0275.99——
NVFP4Model=Llama2-7B, A-W Quant Type=NVFP42026.0275.02—-0.97
HiF4Model=Llama2-7B, A-W Quant Type=HiF42026.0275—-0.99
NVFP4+PTSModel=Llama2-7B, A-W Quant Type=NVFP4+PTS2026.0274.92—-1.07
HiF4+HiGPTQModel=Llama2-7B, A-W Quant Type=HiF4+HiGPTQ2026.0274.91—-1.08
Mamba2-2.7bBit-Flips=9, Attack=COBRA2025.1262.691.33—
Mamba-2.8bBit-Flips=1, Attack=COBRA2025.1262.5926—
Quamba-2.8b-w8a8Bit-Flips=1, Attack=COBRA2025.1262.060—
Quamba-2.8b-w4a16Bit-Flips=1, Attack=COBRA2025.1261.660—
Mamba2-1.3bBit-Flips=5, Attack=COBRA2025.1256.690—
Mamba-1.4bBit-Flips=1, Attack=COBRA2025.1255.760—
Quamba-1.4b-w4a16Bit-Flips=1, Attack=COBRA2025.1255.170—
Hymba-1.5bBit-Flips=2, Attack=COBRA2025.1253.550.67—
Quamba-1.4b-w8a8Bit-Flips=4, Attack=COBRA2025.1249.3922.66—
Mamba2-370mBit-Flips=4, Attack=COBRA2025.1243.088—
Mamba-370mBit-Flips=4, Attack=COBRA2025.1243.0422—
NVFP4Model=Mistral-7B, A-W Quant Type=NVFP42026.0226.69—-53.81