Model Merging on GLUE (CoLA, MRPC, RTE, SST-2)
75.9Absolute AccuracyTask arithmetic
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Task arithmeticFine-tuning configuration=Original, Backbone=Llama2-7b2025.08 | 75.9 | 91.7 | |
| TSVM + DisTaCFine-tuning configuration=Low Confidence, Backbone=RoBERTa-l2025.08 | 75.8 | 90.8 | |
| Task arithmeticFine-tuning configuration=Low Confidence, Backbone=Llama2-7b2025.08 | 75.7 | 95.1 | |
| Task arithmetic + DisTaCFine-tuning configuration=Norm Mismatch, Backbone=Llama2-7b2025.08 | 75 | 91.1 | |
| Consensus-mergingFine-tuning configuration=Original, Backbone=Llama2-7b2025.08 | 74.5 | 89.7 | |
| TSVMFine-tuning configuration=Low Confidence, Backbone=RoBERTa-l2025.08 | 73.3 | 85.6 | |
| Task arithmetic + DisTaCFine-tuning configuration=Low Confidence, Backbone=Llama2-7b2025.08 | 73 | 95.9 | |
| Ties-merging + DisTaCFine-tuning configuration=Low Confidence, Backbone=RoBERTa-l2025.08 | 72.5 | 86.4 | |
| TSVM + DisTaCFine-tuning configuration=Low Confidence, Backbone=Llama2-7b2025.08 | 72.4 | 97.1 | |
| TSVMFine-tuning configuration=Original, Backbone=RoBERTa-l2025.08 | 72 | 87.8 | |
| Consensus-merging + DisTaCFine-tuning configuration=Low Confidence, Backbone=Llama2-7b2025.08 | 72 | 94.9 | |
| Ties-mergingFine-tuning configuration=Low Confidence, Backbone=RoBERTa-l2025.08 | 71.8 | 84 | |
| Task arithmetic + DisTaCFine-tuning configuration=Low Confidence, Backbone=RoBERTa-l2025.08 | 70 | 82.3 | |
| Consensus-merging + DisTaCFine-tuning configuration=Low Confidence, Backbone=RoBERTa-l2025.08 | 69.8 | 82.3 | |
| TSVMFine-tuning configuration=Low Confidence, Backbone=RoBERTa-b2025.08 | 69.6 | 84.3 | |
| Consensus-merging + DisTaCFine-tuning configuration=Norm Mismatch, Backbone=Llama2-7b2025.08 | 68.4 | 82.3 | |
| Task arithmeticFine-tuning configuration=Original, Backbone=RoBERTa-l2025.08 | 68.3 | 82.4 | |
| TSVMFine-tuning configuration=Low Confidence, Backbone=Llama2-7b2025.08 | 68.1 | 84.6 | |
| Consensus-mergingFine-tuning configuration=Original, Backbone=RoBERTa-l2025.08 | 67.9 | 81.4 | |
| TSVM + DisTaCFine-tuning configuration=Low Confidence, Backbone=RoBERTa-b2025.08 | 67.5 | 82.4 | |
| TSVM + DisTaCFine-tuning configuration=Norm Mismatch, Backbone=RoBERTa-l2025.08 | 66.3 | 84.1 | |
| TSVMFine-tuning configuration=Original, Backbone=Llama2-7b2025.08 | 66.1 | 78.5 | |
| TSVMFine-tuning configuration=Original, Backbone=RoBERTa-b2025.08 | 65.8 | 80.8 | |
| Ties-mergingFine-tuning configuration=Original, Backbone=RoBERTa-l2025.08 | 65.7 | 80.7 | |
| Ties-mergingFine-tuning configuration=Low Confidence, Backbone=RoBERTa-b2025.08 | 65.4 | 79.1 | |
| Consensus-mergingFine-tuning configuration=Low Confidence, Backbone=RoBERTa-l2025.08 | 65.2 | 75.5 | |
| TSVM + DisTaCFine-tuning configuration=Norm Mismatch, Backbone=RoBERTa-b2025.08 | 65.1 | 79.6 | |
| Consensus-mergingFine-tuning configuration=Low Confidence, Backbone=Llama2-7b2025.08 | 65 | 79.1 | |
| TSVM + DisTaCFine-tuning configuration=Norm Mismatch, Backbone=Llama2-7b2025.08 | 64.6 | 77.4 | |
| Task arithmeticFine-tuning configuration=Low Confidence, Backbone=RoBERTa-l2025.08 | 64.5 | 73.9 | |
| Task arithmetic + DisTaCFine-tuning configuration=Norm Mismatch, Backbone=RoBERTa-l2025.08 | 64.4 | 80.5 | |
| Ties-merging + DisTaCFine-tuning configuration=Low Confidence, Backbone=RoBERTa-b2025.08 | 64.4 | 78 | |
| Consensus-merging + DisTaCFine-tuning configuration=Norm Mismatch, Backbone=RoBERTa-l2025.08 | 63.4 | 79 | |
| Task arithmetic + DisTaCFine-tuning configuration=Low Confidence, Backbone=RoBERTa-b2025.08 | 62.5 | 74.6 | |
| Ties-merging + DisTaCFine-tuning configuration=Norm Mismatch, Backbone=RoBERTa-b2025.08 | 62.4 | 76.4 | |
| Consensus-merging + DisTaCFine-tuning configuration=Low Confidence, Backbone=RoBERTa-b2025.08 | 62.2 | 74.3 | |
| Task arithmeticFine-tuning configuration=Low Confidence, Backbone=RoBERTa-b2025.08 | 61.3 | 72.6 | |
| Consensus-mergingFine-tuning configuration=Original, Backbone=RoBERTa-b2025.08 | 61.3 | 73.7 | |
| Consensus-mergingFine-tuning configuration=Low Confidence, Backbone=RoBERTa-b2025.08 | 61.2 | 72.3 | |
| Task arithmeticFine-tuning configuration=Original, Backbone=RoBERTa-b2025.08 | 60.9 | 73.5 | |
| Ties-mergingFine-tuning configuration=Original, Backbone=RoBERTa-b2025.08 | 60.9 | 74.8 | |
| Consensus-merging + DisTaCFine-tuning configuration=Norm Mismatch, Backbone=RoBERTa-b2025.08 | 60.5 | 72.5 | |
| Task arithmetic + DisTaCFine-tuning configuration=Norm Mismatch, Backbone=RoBERTa-b2025.08 | 59.9 | 71.7 | |
| Ties-merging + DisTaCFine-tuning configuration=Norm Mismatch, Backbone=RoBERTa-l2025.08 | 59.4 | 75.9 | |
| Ties-merging + DisTaCFine-tuning configuration=Low Confidence, Backbone=Llama2-7b2025.08 | 58.9 | 78.4 | |
| TSVMFine-tuning configuration=Norm Mismatch, Backbone=RoBERTa-b2025.08 | 58.8 | 71.1 | |
| Ties-mergingFine-tuning configuration=Original, Backbone=Llama2-7b2025.08 | 58.3 | 80.7 | |
| Consensus-mergingFine-tuning configuration=Norm Mismatch, Backbone=Llama2-7b2025.08 | 58.3 | 68.6 | |
| Consensus-mergingFine-tuning configuration=Norm Mismatch, Backbone=RoBERTa-b2025.08 | 58.1 | 70 | |
| Task arithmeticFine-tuning configuration=Norm Mismatch, Backbone=RoBERTa-b2025.08 | 56.8 | 68.5 | |
| TSVMFine-tuning configuration=Norm Mismatch, Backbone=Llama2-7b2025.08 | 55.5 | 65.5 | |
| Task arithmeticFine-tuning configuration=Norm Mismatch, Backbone=Llama2-7b2025.08 | 55.3 | 64.7 | |
| TSVMFine-tuning configuration=Norm Mismatch, Backbone=RoBERTa-l2025.08 | 48 | 60.7 | |
| Task arithmeticFine-tuning configuration=Norm Mismatch, Backbone=RoBERTa-l2025.08 | 46 | 58.1 | |
| Ties-merging + DisTaCFine-tuning configuration=Norm Mismatch, Backbone=Llama2-7b2025.08 | 44 | 51.6 | |
| Ties-mergingFine-tuning configuration=Norm Mismatch, Backbone=RoBERTa-l2025.08 | 40.8 | 51.3 | |
| Ties-mergingFine-tuning configuration=Norm Mismatch, Backbone=Llama2-7b2025.08 | 40.6 | 47.7 | |
| Ties-mergingFine-tuning configuration=Norm Mismatch, Backbone=RoBERTa-b2025.08 | 39.9 | 46.1 | |
| Ties-mergingFine-tuning configuration=Low Confidence, Backbone=Llama2-7b2025.08 | 38.3 | 47.5 | |
| Consensus-mergingFine-tuning configuration=Norm Mismatch, Backbone=RoBERTa-l2025.08 | 38.1 | 47.3 |