Instruction Tuning Data Selection Evaluation on ADA 10% selected-token budget
71.6Task ScoreQuality
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| QualitySelected-token budget=10%, Number of seeds=3, Base model=fixed, SFT recipe=fixed2026.07 | 71.6 | 56.2 | 85.1 | 14.4 | 207 | 42.8 | 26.8 | 36 | |
| AASSelected-token budget=10%, Number of seeds=3, Base model=fixed, SFT recipe=fixed2026.07 | 71.4 | 56.1 | 81.9 | 10.5 | 190 | 42.9 | 25.2 | 13 | |
| LossSelected-token budget=10%, Number of seeds=3, Base model=fixed, SFT recipe=fixed2026.07 | 70.8 | 54.8 | 76.9 | 11.3 | 202 | 40.2 | 28.5 | 42 | |
| DiversitySelected-token budget=10%, Number of seeds=3, Base model=fixed, SFT recipe=fixed2026.07 | 69.4 | 52.1 | 82.3 | 10.4 | 189 | 42.6 | 25.4 | 19 | |
| RandomSelected-token budget=10%, Number of seeds=3, Base model=fixed, SFT recipe=fixed2026.07 | 67.9 | 50 | 81.2 | 9.6 | 184 | 42.1 | 24.7 | 0 |