Supervised Fine-Tuning on MMMLU Marathi (test)
34.6MMMLU AccuracyTV + KMM
Evaluation results
| Method | Links | ||
|---|---|---|---|
| TV + KMMk=25, Pre-trained model=Qwen3-1.7b2026.05 | 34.6 | 33.5 | |
| One Step + KMMk=25, Pre-trained model=Qwen3-1.7b2026.05 | 34.58 | 31 | |
| Randomk=10, Pre-trained model=Qwen3-1.7b2026.05 | 34.54 | 21.5 | |
| DataModelk=10, Pre-trained model=Qwen3-1.7b2026.05 | 34.52 | 33 | |
| TVk=5, Pre-trained model=Qwen3-1.7b2026.05 | 34.49 | 17.5 | |
| One Stepk=15, Pre-trained model=Qwen3-1.7b2026.05 | 34.47 | 19 | |
| GradEX-REk=25, Pre-trained model=Qwen3-1.7b2026.05 | 34.47 | 11.5 | |
| GradEX-FSk=20, Pre-trained model=Qwen3-1.7b2026.05 | 34.42 | 13 | |
| DataModel-CSk=5, Pre-trained model=Qwen3-1.7b2026.05 | 34.03 | 0 |