Mathematical Reasoning on MGSM Monolingual P, I, Q - (X) (Avg)
100Language ConsistencyBloom-7.1B (Full scope SFT)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Bloom-7.1B (Full scope SFT)Model=Bloom-7.1B, Fine-tuning Strategy=Full scope SFT, # Trainable Param=7.1B2026.01 | 100 | 1.47 | |
| Bloom-7.1B (Selective SFT)Model=Bloom-7.1B, Fine-tuning Strategy=Selective SFT, # Trainable Param=0.5B2026.01 | 100 | 3.6 | |
| Qwen-3-32B (Full scope SFT)Model=Qwen-3-32B, Fine-tuning Strategy=Full scope SFT, # Trainable Param=32B2026.01 | 99.47 | 90.53 | |
| Qwen-3-32B (Selective SFT)Model=Qwen-3-32B, Fine-tuning Strategy=Selective SFT, # Trainable Param=1.5B2026.01 | 99.2 | 86.8 | |
| Bloom-7.1B (Random Selective SFT)Model=Bloom-7.1B, Fine-tuning Strategy=Random Selective SFT, # Trainable Param=0.5B2026.01 | 69.47 | 0 | |
| Qwen-3-32B (Random Selective SFT)Model=Qwen-3-32B, Fine-tuning Strategy=Random Selective SFT, # Trainable Param=1.5B2026.01 | 65.87 | 0.13 | |
| Qwen-3-32B (Pre-Finetuning)Model=Qwen-3-32B, Fine-tuning Strategy=Pre-Finetuning, # Trainable Param=NA2026.01 | 65.56 | 66.6 | |
| Bloom-7.1B (Pre-Finetuning)Model=Bloom-7.1B, Fine-tuning Strategy=Pre-Finetuning, # Trainable Param=NA2026.01 | 34 | 0.67 |