Linear Mode Connectivity on WikiText
0.18Loss BarrierDual Learned Matching
Evaluation Results
| Method | Links | |
|---|---|---|
| Dual Learned MatchingModel Size=160M2026.06 | 0.18 | |
| Dual Learned MatchingModel Size=1B2026.06 | 0.23 | |
| Enhanced Learned MatchingModel Size=160M2026.06 | 0.27 | |
| Enhanced Learned MatchingModel Size=1B2026.06 | 0.28 | |
| Dual Learned MatchingModel Size=410M2026.06 | 0.37 | |
| Learned MatchingModel Size=160M2026.06 | 0.39 | |
| Learned MatchingModel Size=1B2026.06 | 0.42 | |
| Enhanced Learned MatchingModel Size=410M2026.06 | 0.44 | |
| Learned MatchingModel Size=410M2026.06 | 0.57 | |
| Dual Learned MatchingModel Size=14M2026.06 | 0.86 | |
| Enhanced Learned MatchingModel Size=14M2026.06 | 1 | |
| Learned MatchingModel Size=14M2026.06 | 1.26 | |
| Raw InterpolationModel Size=1B2026.06 | 4.6 | |
| Weight MatchingModel Size=1B2026.06 | 4.75 | |
| Raw InterpolationModel Size=410M2026.06 | 4.99 | |
| Weight MatchingModel Size=410M2026.06 | 5.65 | |
| Weight MatchingModel Size=160M2026.06 | 6.04 | |
| Raw InterpolationModel Size=160M2026.06 | 6.28 | |
| Weight MatchingModel Size=14M2026.06 | 8.17 | |
| Raw InterpolationModel Size=14M2026.06 | 9.89 |