List your company
ResearchFormerly WizWand

Language Modeling on WikiText (Adversarial Robustness Evaluation)

1.7PPL Change (%)LoRA
-0.00811.52123.0534.579Apr 6, 2026
Updated 6mo ago

Evaluation results

MethodLinks
LoRABackbone=Mistral-7B, Learning Rate=10⁻³, Tokens per domain=200K, Training Stage=+Fin2026.041.7——
LoRABackbone=Mistral-7B, Learning Rate=10⁻³, Tokens per domain=200K, Training Stage=+PG2026.045.7——
LoRABackbone=Mistral-7B, Learning Rate=10⁻³, Tokens per domain=200K, Training Stage=+Med2026.0410.8——
LoRABackbone=Mistral-7B, Learning Rate=10⁻³, Tokens per domain=200K, Training Stage=+MdQ2026.0410.9——
LoRABackbone=Mistral-7B, Learning Rate=10⁻³, Tokens per domain=200K, Training Stage=+LB2026.0415.3——
GAINBackbone=Mistral-7B, Learning Rate=10⁻³, Tokens per domain=200K, Training Stage=+HS2026.0422.1——
GAINBackbone=Mistral-7B, Learning Rate=10⁻³, Tokens per domain=200K, Training Stage=+SST2026.0422.4——
GAINBackbone=Mistral-7B, Learning Rate=10⁻³, Tokens per domain=200K, Training Stage=+LB2026.0422.6——
GAINBackbone=Mistral-7B, Learning Rate=10⁻³, Tokens per domain=200K, Training Stage=+MdQ2026.0422.7——
GAINBackbone=Mistral-7B, Learning Rate=10⁻³, Tokens per domain=200K, Training Stage=+PG2026.0422.9——
GAINBackbone=Mistral-7B, Learning Rate=10⁻³, Tokens per domain=200K, Training Stage=+WT2026.0423.1——
GAINBackbone=Mistral-7B, Learning Rate=10⁻³, Tokens per domain=200K, Training Stage=+Fin2026.0423.1——
GAINBackbone=Mistral-7B, Learning Rate=10⁻³, Tokens per domain=200K, Training Stage=+Med2026.0423.2——
LoRABackbone=Mistral-7B, Learning Rate=10⁻³, Tokens per domain=200K, Training Stage=+WT2026.0430.9——
LoRABackbone=Mistral-7B, Learning Rate=10⁻³, Tokens per domain=200K, Training Stage=+SST2026.0443.6——
LoRABackbone=Mistral-7B, Learning Rate=10⁻³, Tokens per domain=200K, Training Stage=+HS2026.0444.4——
Hymba-1.5bBit-Flips=2, Attack=COBRA2025.12—14.417,620
Mamba-1.4bBit-Flips=1, Attack=COBRA2025.12—18.943,750,000
Mamba-2.8bBit-Flips=1, Attack=COBRA2025.12—16.31123,000
Mamba-370mBit-Flips=4, Attack=COBRA2025.12—24.87126,000,000,000
Mamba2-1.3bBit-Flips=5, Attack=COBRA2025.12—18.762,410,000
Mamba2-2.7bBit-Flips=9, Attack=COBRA2025.12—16.5924,800
Mamba2-370mBit-Flips=4, Attack=COBRA2025.12—26.6924,800
Quamba-1.4b-w4a16Bit-Flips=1, Attack=COBRA2025.12—23.62606.673
Quamba-1.4b-w8a8Bit-Flips=4, Attack=COBRA2025.12—22.19350.08
Quamba-2.8b-w4a16Bit-Flips=1, Attack=COBRA2025.12—9.91765,000
Quamba-2.8b-w8a8Bit-Flips=1, Attack=COBRA2025.12—19.5312,600,000,000