Proxy Consistency Evaluation on 50B tokens corpus
0.66Spearman Rho (General)DeMix
Evaluation results
| Method | Links | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| DeMixTotal Budget (B)=351, Pre-Cost=50×7, No. Proxies=112, Budget / Proxy=0.012026.01 | 0.66 | 0.76 | 0.97 | 0.8 | 0.2 | 0.37 | 0.93 | 0.5 | 101 | 82 | 73 | 85 | |
| DeMixTotal Budget (B)=211, Pre-Cost=30×7, No. Proxies=112, Budget / Proxy=0.012026.01 | 0.64 | 0.81 | 0.98 | 0.81 | 0 | 0.8 | 0.97 | 0.59 | 100 | 81 | 68 | 83 | |
| Trained Proxy (RegMix/CLIMB)Total Budget (B)=1344, Pre-Cost=-, No. Proxies=112, Budget / Proxy=122026.01 | 0.54 | 0.94 | 0.97 | 0.82 | 0.03 | 0.7 | 0.97 | 0.57 | 99 | 89 | 74 | 87 | |
| Trained Proxy (RegMix/CLIMB)Total Budget (B)=896, Pre-Cost=-, No. Proxies=112, Budget / Proxy=82026.01 | 0.41 | 0.77 | 0.97 | 0.71 | 0.4 | 0.43 | 0.57 | 0.47 | 99 | 83 | 69 | 83 | |
| DeMixTotal Budget (B)=15, Pre-Cost=2×7, No. Proxies=112, Budget / Proxy=0.012026.01 | 0.25 | 0.51 | 0.88 | 0.55 | -0.17 | 0.01 | 0.97 | 0.27 | 98 | 77 | 54 | 76 | |
| DeMixTotal Budget (B)=71, Pre-Cost=10×7, No. Proxies=112, Budget / Proxy=0.012026.01 | 0.13 | 0.74 | 0.95 | 0.6 | -0.1 | 0.63 | 0.71 | 0.41 | 99 | 79 | 64 | 80 | |
| Trained Proxy (RegMix/CLIMB)Total Budget (B)=224, Pre-Cost=-, No. Proxies=112, Budget / Proxy=22026.01 | 0.12 | 0.6 | 0.85 | 0.53 | 0.17 | 0.63 | -0.2 | 0.2 | 98 | 76 | 57 | 77 | |
| Trained Proxy (RegMix/CLIMB)Total Budget (B)=448, Pre-Cost=-, No. Proxies=112, Budget / Proxy=42026.01 | -0.24 | 0.81 | 0.96 | 0.51 | -0.27 | 0.67 | 0.73 | 0.38 | 98 | 80 | 63 | 80 |