Tabular Data Synthesis on Average of 5 Datasets (Adult, Shoppers, Beijing, and two others)
95.47CDETabGRAA (base)
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| TabGRAA (base)Alignment Framework=GRAA, Loss Function Variant=base2026.04 | 95.47 | 58.8 | 96.73 | 50.06 | 35.01 | 67.95 | |
| TabDPO (base)Alignment Framework=DPO, Loss Function Variant=base2026.04 | 95.02 | 57.58 | 95.81 | 49.87 | 33.4 | 68.34 | |
| TabDPO + Gradient diff.Alignment Framework=DPO, Gradient Difference=true2026.04 | 95 | 56.46 | 96.75 | 49.83 | 33.32 | 68.91 | |
| TabGRAA + LogsigmoidAlignment Framework=GRAA, Logsigmoid=true2026.04 | 94.86 | 59.26 | 97.74 | 49.6 | 34.32 | 69.02 | |
| TabGRAA + Logs. + Grad. diffAlignment Framework=GRAA, Logsigmoid=true, Gradient Difference=true2026.04 | 94.84 | 57.35 | 97.04 | 50.03 | 34.19 | 68.73 | |
| TabDPO + KL penaltyAlignment Framework=DPO, KL Penalty=true2026.04 | 94.74 | 56.93 | 95.5 | 49.39 | 33.35 | 68.5 | |
| TabNPO + Gradient diff.Alignment Framework=NPO, Gradient Difference=true2026.04 | 88.26 | 54.74 | 85.52 | 49.47 | 15.38 | 77.77 | |
| TabNPO + KL penaltyAlignment Framework=NPO, KL Penalty=true2026.04 | 87.94 | 54.55 | 85.4 | 49.11 | 15.1 | 77.75 | |
| TabNPO (base)Alignment Framework=NPO, Loss Function Variant=base2026.04 | 87.9 | 56.58 | 85.8 | 49.28 | 14.84 | 77.94 | |
| GReaT-FT+Alignment Framework=Baseline2026.04 | 87.42 | 58.63 | 85.82 | 46.53 | 33.4 | 81.51 | |
| TabKTO (base)Alignment Framework=KTO, Loss Function Variant=base2026.04 | 87.09 | 56.44 | 85.14 | 49.19 | 14.53 | 78.24 | |
| TabKTO + LogsigmoidAlignment Framework=KTO, Logsigmoid=true2026.04 | 87.09 | 55.64 | 84.73 | 48.93 | 14.72 | 78.34 | |
| TabKTO + Logs. + Grad. diffAlignment Framework=KTO, Logsigmoid=true, Gradient Difference=true2026.04 | 87.09 | 55.27 | 84.72 | 48.95 | 14.61 | 78.35 | |
| GReaTAlignment Framework=Baseline2026.04 | 86.86 | 58.72 | 83.41 | 47.05 | 32.75 | 81.79 |