Natural Language Understanding on GLUE (Full Task Set)
69.86CoLA AccuracyScaLoRA
Evaluation results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| ScaLoRABackbone=DeBERTaV3-base, Number of parameters=184M, Rank (r)=4, Scaling coefficient=8, Number of random runs=32025.10 | 69.86 | 90.36 | 90.28 | 94.34 | 92.1 | 87.61 | 95.83 | 88.98 | 91.47 | |
| Weight AveragingMethod Category=Data-free Methods, Architecture=Flan-T5-base, Fine-tuning Protocol=LoRA fine-tuned2025.11 | 69.7 | 59.7 | 78.9 | 90.1 | 83.8 | 80.5 | 91.2 | 78.2 | — | |
| MoRABackbone=DeBERTaV3-base, Number of parameters=184M, Rank (r)=4, Scaling coefficient=8, Number of random runs=32025.10 | 69.67 | 90.05 | 89.62 | 93.81 | 91.83 | 85.44 | 95.45 | 88.35 | 90.9 | |
| MDA TAMethod Category=Data-free Methods, Architecture=Flan-T5-base, Fine-tuning Protocol=LoRA fine-tuned2025.11 | 69.5 | 76.9 | 76.9 | 89.6 | 83.9 | 82.4 | 92.5 | 82.1 | — | |
| TSV TAMethod Category=Data-free Methods, Architecture=Flan-T5-base, Fine-tuning Protocol=LoRA fine-tuned2025.11 | 69.3 | 77.1 | 80.4 | 90 | 83.6 | 79.1 | 92.5 | 81.8 | — | |
| Full FTBackbone=DeBERTaV3-base, Number of parameters=184M2025.10 | 69.19 | 89.9 | 89.46 | 94.03 | 92.4 | 83.75 | 95.63 | 88.25 | 91.6 | |
| IndividualMethod Category=Non-merging Methods, Architecture=Flan-T5-base, Fine-tuning Protocol=LoRA fine-tuned2025.11 | 69.1 | 82.7 | 85.5 | 90.9 | 84 | 84.4 | 92.9 | 84.6 | — | |
| Concrete TAMethod Category=Data-free Methods, Architecture=Flan-T5-base, Fine-tuning Protocol=LoRA fine-tuned2025.11 | 69.1 | 58.1 | 78.4 | 89.9 | 83.5 | 79.4 | 91.6 | 78 | — | |
| ISO TAMethod Category=Data-free Methods, Architecture=Flan-T5-base, Fine-tuning Protocol=LoRA fine-tuned2025.11 | 69.1 | 57.4 | 76.7 | 88.6 | 82.7 | 80.1 | 91.3 | 76.2 | — | |
| DoGE TAMethod Category=Data-free Methods, Architecture=Flan-T5-base, Fine-tuning Protocol=LoRA fine-tuned2025.11 | 69.1 | 71.9 | 80.9 | 90.3 | 83.5 | 79.8 | 92.5 | 79.9 | — | |
| AdaMerging++Method Category=Data-based Optimization Methods, Architecture=Flan-T5-base, Fine-tuning Protocol=LoRA fine-tuned2025.11 | 69.1 | 60.3 | 78.4 | 90 | 83.6 | 79.1 | 91.6 | 78.3 | — | |
| Concrete AMMethod Category=Data-based Optimization Methods, Architecture=Flan-T5-base, Fine-tuning Protocol=LoRA fine-tuned2025.11 | 69 | 59.4 | 80.1 | 89.9 | 82.9 | 79.1 | 91.7 | 78.5 | — | |
| HiRABackbone=DeBERTaV3-base, Number of parameters=184M, Rank (r)=4, Scaling coefficient=8, Number of random runs=32025.10 | 68.82 | 90.24 | 89.95 | 94.15 | 92.19 | 85.68 | 95.53 | 88.46 | 91.15 | |
| Task ArithmeticMethod Category=Data-free Methods, Architecture=Flan-T5-base, Fine-tuning Protocol=LoRA fine-tuned2025.11 | 68.8 | 55.2 | 78.7 | 89.8 | 83.7 | 79.1 | 91.5 | 77.4 | — | |
| Ties-MergingMethod Category=Data-free Methods, Architecture=Flan-T5-base, Fine-tuning Protocol=LoRA fine-tuned2025.11 | 68.3 | 56.3 | 79.2 | 89.8 | 83.7 | 79.4 | 91.6 | 77.5 | — | |
| LoRABackbone=DeBERTaV3-base, Number of parameters=184M, Rank (r)=4, Scaling coefficient=8, Number of random runs=32025.10 | 68.1 | 90.25 | 89.46 | 94.3 | 91.86 | 84.48 | 95.49 | 88.13 | 91.09 |