List your company

Language Modeling on Qwen3-0.6B (val)

31.45Validation PerplexityDelta Block
31.412831.663931.91532.1661May 13, 2026

Evaluation results

MethodLinks
Delta BlockL (Layers)=28, N (Routing layers)=28, Steps=10K, Learning Rate (lr)=6×10−42026.0531.4511725.2
BaselineL (Layers)=28, Steps=10K, Learning Rate (lr)=6×10−42026.0532.2217313.8
AttnResL (Layers)=28, N (Routing layers)=28, Steps=10K, Learning Rate (lr)=6×10−42026.0532.3811128.4