Code-switching Ratio Calculation on Multilingual Query Set
1.16CS Ratio (zh)SFT (Baseline)
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| SFT (Baseline)Model=Llama-3.1-8B, Training Data Size=210k2025.07 | 1.16 | 0.67 | 0.57 | |
| SFT+PenaltyModel=Gemma-2-9B, Training Data Size=210k2025.07 | 0.96 | 0.11 | 0.63 | |
| SFT (Baseline)Model=Qwen3-8B-Base, Training Data Size=210k2025.07 | 0.96 | 0.16 | 0.43 | |
| SFT+GRPOModel=Qwen3-8B-Base, Training Data Size=210k2025.07 | 0.83 | 0.09 | 0.32 | |
| SFT (Baseline)Model=Qwen3-8B-Base, Training Data Size=110k2025.07 | 0.83 | 0.17 | 0.25 | |
| SFT (Baseline)Model=Qwen3-1.7B-Base, Training Data Size=210k2025.07 | 0.81 | 0.19 | 0.36 | |
| SFT (Baseline)Model=Gemma-2-9B, Training Data Size=110k2025.07 | 0.8 | 0.05 | 0.5 | |
| SFT (Baseline)Model=Gemma-2-9B, Training Data Size=210k2025.07 | 0.78 | 0.12 | 0.81 | |
| SFT+GRPOModel=Gemma-2-9B, Training Data Size=210k2025.07 | 0.78 | 0.16 | 0.7 | |
| SFT+GRPOModel=Gemma-2-9B, Training Data Size=110k2025.07 | 0.78 | 0.04 | 0.48 | |
| SFT+PenaltyModel=Gemma-2-9B, Training Data Size=110k2025.07 | 0.78 | 0.07 | 0.29 | |
| SFT+PenaltyModel=Qwen3-8B-Base, Training Data Size=110k2025.07 | 0.76 | 0.08 | 0.18 | |
| SFT (Baseline)Model=Gemma-2-2B, Training Data Size=210k2025.07 | 0.74 | 0.57 | 3.45 | |
| SFT+GRPOModel=Gemma-2-2B, Training Data Size=210k2025.07 | 0.74 | 0.49 | 3.44 | |
| SFT+GRPOModel=Gemma-2-2B, Training Data Size=110k2025.07 | 0.72 | 0.36 | 1.08 | |
| SFT+PenaltyModel=Qwen3-8B-Base, Training Data Size=210k2025.07 | 0.7 | 0.12 | 0.23 | |
| SFT (Baseline)Model=Gemma-2-2B, Training Data Size=110k2025.07 | 0.68 | 0.42 | 1.06 | |
| SFT (Baseline)Model=Qwen3-1.7B-Base, Training Data Size=110k2025.07 | 0.68 | 0.19 | 0.23 | |
| SFT+GRPOModel=Qwen3-1.7B-Base, Training Data Size=110k2025.07 | 0.68 | 0.19 | 0.2 | |
| SFT+PenaltyModel=Gemma-2-2B, Training Data Size=210k2025.07 | 0.67 | 0.41 | 1.18 | |
| SFT+GRPOModel=Qwen3-8B-Base, Training Data Size=110k2025.07 | 0.67 | 0.06 | 0.12 | |
| SFT+PenaltyModel=Gemma-2-2B, Training Data Size=110k2025.07 | 0.66 | 0.32 | 0.7 | |
| SFT+GRPOModel=Qwen3-1.7B-Base, Training Data Size=210k2025.07 | 0.66 | 0.11 | 0.34 | |
| SASFTModel=Qwen3-8B-Base, Training Data Size=210k2025.07 | 0.66 | 0.07 | 0.07 | |
| SASFTModel=Qwen3-8B-Base, Training Data Size=110k2025.07 | 0.62 | 0.07 | 0.05 | |
| SFT+PenaltyModel=Qwen3-1.7B-Base, Training Data Size=210k2025.07 | 0.53 | 0.09 | 0.06 | |
| SFT+GRPOModel=Llama-3.1-8B, Training Data Size=210k2025.07 | 0.52 | 0.3 | 0.37 | |
| SFT+PenaltyModel=Qwen3-1.7B-Base, Training Data Size=110k2025.07 | 0.49 | 0.07 | 0.06 | |
| SFT (Baseline)Model=Llama-3.1-8B, Training Data Size=110k2025.07 | 0.48 | 0.79 | 0.37 | |
| SASFTModel=Gemma-2-2B, Training Data Size=210k2025.07 | 0.42 | 0.22 | 0.73 | |
| SFT+PenaltyModel=Llama-3.1-8B, Training Data Size=210k2025.07 | 0.42 | 0.33 | 0.54 | |
| SASFTModel=Gemma-2-9B, Training Data Size=210k2025.07 | 0.41 | 0.01 | 0.13 | |
| SFT+GRPOModel=Llama-3.1-8B, Training Data Size=110k2025.07 | 0.4 | 0.34 | 0.95 | |
| SFT+PenaltyModel=Llama-3.1-8B, Training Data Size=110k2025.07 | 0.39 | 0.27 | 0.31 | |
| SASFTModel=Gemma-2-9B, Training Data Size=110k2025.07 | 0.34 | 0.01 | 0.24 | |
| SASFTModel=Gemma-2-2B, Training Data Size=110k2025.07 | 0.32 | 0.16 | 0.34 | |
| SASFTModel=Llama-3.1-8B, Training Data Size=110k2025.07 | 0.32 | 0.22 | 0.2 | |
| SASFTModel=Qwen3-1.7B-Base, Training Data Size=110k2025.07 | 0.31 | 0.03 | 0.02 | |
| SASFTModel=Llama-3.1-8B, Training Data Size=210k2025.07 | 0.28 | 0.28 | 0.46 | |
| SASFTModel=Qwen3-1.7B-Base, Training Data Size=210k2025.07 | 0.22 | 0.03 | 0 |