List your company

Language Modeling on LM Evaluation Harness (test)

74.11WG (Winograd Schema)Llama3.1-Mamba2-DPO
57.93862.136566.33570.5335Aug 27, 2024Nov 19, 2024Feb 11, 2025May 6, 2025Jul 29, 2025Oct 21, 2025Jan 14, 2026

Evaluation results

MethodLinks
Llama3.1-Mamba2-DPOpercentage_attention_layers=50%, evaluation_mode=zero-shot2024.0874.1180.0379.6984.8159.7359.744450.2274.646.1265.31—————
Llama-3.1-8B-Instructevaluation_mode=zero-shot2024.0873.8880.7979.2181.7855.268.1243.242.6775.244.7864.48—————
Llama3.1-Mamba-DPOpercentage_attention_layers=50%, evaluation_mode=zero-shot2024.0873.880.4177.3684.0156.5763.544.246.0774.443.4464.38—————
Llama3.1-Mambapercentage_attention_layers=50%, evaluation_mode=zero-shot2024.0872.7779.3375.9182.2453.8462.1342.840.027242.1162.32—————
Llama3.1-Mamba2percentage_attention_layers=50%, evaluation_mode=zero-shot2024.0871.7478.8975.3682.252.6561.0141.640.3172.642.1161.85—————
Mamba2-Llama3percentage_attention_layers=50%, evaluation_mode=zero-shot2024.0871.5181.4579.4778.8358.1955.744.257.7472.438.8563.84—————
TRI Mamba-7Bevaluation_mode=zero-shot2024.0871.4281.0177.9377.5346.6733.3946.232.0972.337.9957.65—————
Nvidia Hybrid Mamba-8Bevaluation_mode=zero-shot2024.0871.2779.6577.6877.2347.751.4642.838.7269.839.7159.6—————
Mamba-Llama3percentage_attention_layers=50%, evaluation_mode=zero-shot2024.0868.9878.0278.4374.4551.9657.814447.697338.5661.3—————
Mamba-Zephyrpercentage_attention_layers=50%, evaluation_mode=zero-shot2024.0868.8280.3676.9181.455.6355.4342.641.9972.642.261.79—————
Llama-3.2-3B-Instructevaluation_mode=zero-shot2024.0867.4875.6870.4374.0745.960.433638.0169.640.6757.83—————
Llama3.2-Mamba-DPOpercentage_attention_layers=50%, evaluation_mode=zero-shot2024.0867.477.3172.5679.9752.6555.0941.648.537043.6460.88—————
Llama3.2-Mambapercentage_attention_layers=50%, evaluation_mode=zero-shot2024.0867.3277.3170.3777.6548.3854.4839.442.0266.440.2958.36—————
Llama3.2-Mamba2-DPOpercentage_attention_layers=50%, evaluation_mode=zero-shot2024.0867.3277.6974.4580.2654.152.4742.450.2865.443.4460.78—————
Llama3.2-Mamba2percentage_attention_layers=50%, evaluation_mode=zero-shot2024.0866.0676.0169.1376.6846.6753.1238.834.7863.839.8156.49—————
Mamba2-Llama3percentage_attention_layers=25%, evaluation_mode=zero-shot2024.0864.878.7377.776.3552.4753.7142.455.3364.839.2360.55—————
Qwen3-1.7BAttention Mechanism=Global-Attn2026.0164.4875.6866.48——60.48————71———62.9595.9
ROSA-TuningAttention Mechanism=Window-Attn + ROSA, Base Model=Qwen3-1.7B2026.0163.9375.1965.58——60.33————70.5———62.5695.4
Mamba2-Llama3percentage_attention_layers=12.5%, evaluation_mode=zero-shot2024.0863.3876.8273.1475.8450.2650.7839.65065.836.4658.21—————
Mamba-Llama3percentage_attention_layers=25%, evaluation_mode=zero-shot2024.0862.8378.077574.2847.3553.54043.6465.436.9457.7—————
Mamba-Llama3percentage_attention_layers=12.5%, evaluation_mode=zero-shot2024.0859.7575.0871.7170.5843.649.8141.441.4162.434.4555.02—————
Mamba2-Llama3percentage_attention_layers=0%, evaluation_mode=zero-shot2024.0858.5676.8270.7574.1247.9545.193940.262.232.6354.74—————