General Reasoning and Coding on Key-Value Recall Non-Factoid
64.1GSM8K AccuracyREMIX (Random / -)
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| REMIX (Random / -)Backbone=Llama-3-8B, Mixing Strategy (DM1 / DM2)=Random / -2024.11 | 64.1 | 75.9 | 85.3 | 75 | 89.1 | 77.9 | |
| REMIX (K-Pile / -)Backbone=Llama-3-8B, Mixing Strategy (DM1 / DM2)=K-Pile / -2024.11 | 47.3 | 58.4 | 62.2 | 19 | 74.3 | 52.2 | |
| REMIX (Random / K-Pile)Backbone=Llama-3-8B, Mixing Strategy (DM1 / DM2)=Random / K-Pile2024.11 | 45.8 | 45.4 | 74.7 | 51.2 | 86.8 | 60.8 | |
| No MixingBackbone=Llama-3-8B, Mixing Strategy (DM1 / DM2)=No Mixing2024.11 | 24.4 | 27.3 | 49.5 | 26.7 | 66.6 | 38.9 | |
| REMIX (- / Random)Backbone=Llama-3-8B, Mixing Strategy (DM1 / DM2)=- / Random2024.11 | 15.1 | 11.7 | 33.8 | 16.5 | 66.8 | 28.8 | |
| REMIX (- / K-Pile)Backbone=Llama-3-8B, Mixing Strategy (DM1 / DM2)=- / K-Pile2024.11 | 12.8 | 8.8 | 40.5 | 16.8 | 70.2 | 29.8 |