Dataset Diversity Analysis on Post-training LLM Datasets
452.44Vendi ScoreOurs (Provenance-based)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Ours (Provenance-based)Size=570K2026.04 | 452.44 | 0.6385 | |
| OpenHermes-2.5Size=615K2026.04 | 437.76 | 0.6271 | |
| herculesv1Size=463K2026.04 | 397.33 | 0.6121 | |
| tulu-3-sft-mixtureSize=939K2026.04 | 375.78 | 0.6169 | |
| MegaScienceSize=1.2M2026.04 | 373.78 | 0.615 | |
| TextbookReasoningSize=651K2026.04 | 283.75 | 0.5598 | |
| OmniThought-0528Size=301K2026.04 | 162.52 | 0.514 | |
| OpenThoughts3Size=1.2M2026.04 | 133.26 | 0.497 | |
| MiroMind-M1-SFTSize=719K2026.04 | 108.89 | 0.4597 |