Zero-shot Language Understanding and Reasoning on BENCH-PROXY Suite (test)
34.32MMLURandom
Evaluation results
| Method | Links | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| RandomBackbone=GPT-2 XL, Optimizer=Muon, Training update tokens=30B, Data quality subset score=4+5, Evaluation protocol=Zero-shot2026.02 | 34.32 | 33.78 | 46.39 | 68.72 | 39.36 | 47.59 | 50.44 | 32.54 | 29.48 | 36.54 | 41.92 | |
| DSIRBackbone=GPT-2 XL, Optimizer=Muon, Training update tokens=30B, Data quality subset score=4+5, Evaluation protocol=Zero-shot2026.02 | 34.13 | 33.63 | 45.1 | 67.79 | 39.82 | 48.15 | 49.03 | 32.88 | 28.83 | 36.54 | 41.59 | |
| OPUSBackbone=GPT-2 XL, Optimizer=Muon, Training update tokens=30B, Data quality subset score=4+5, Evaluation protocol=Zero-shot2026.02 | 33.83 | 33.64 | 46.3 | 70.67 | 38.95 | 51.14 | 50.62 | 29.15 | 30.47 | 39.42 | 42.42 | |
| RandomBackbone=GPT-2 XL, Optimizer=Muon, Training update tokens=60B, Data quality subset score=4+5, Evaluation protocol=Zero-shot2026.02 | 33.77 | 33.54 | 46.94 | 69.64 | 39.82 | 49.8 | 50.44 | 32.54 | 30.96 | 38.46 | 42.59 | |
| GREATSBackbone=GPT-2 XL, Optimizer=Muon, Training update tokens=30B, Data quality subset score=4+5, Evaluation protocol=Zero-shot2026.02 | 33.58 | 33.02 | 46.32 | 68.93 | 39.61 | 52.57 | 49.21 | 33.9 | 28.01 | 36.54 | 42.17 | |
| UltraFinewebBackbone=GPT-2 XL, Optimizer=Muon, Training update tokens=30B, Data quality subset score=4+5, Evaluation protocol=Zero-shot2026.02 | 33.41 | 33.48 | 44.34 | 68.93 | 38.64 | 48.3 | 49.38 | 33.56 | 29.07 | 36.54 | 41.57 | |
| RandomBackbone=GPT-2 Large, Optimizer=Muon, Training update tokens=60B, Data quality subset score=4+5, Evaluation protocol=Zero-shot2026.02 | 33.21 | 34.03 | 43.66 | 67.95 | 40.07 | 50.04 | 52.56 | 31.86 | 31.61 | 36.54 | 42.15 | |
| DCLM-FastTextBackbone=GPT-2 XL, Optimizer=Muon, Training update tokens=30B, Data quality subset score=4+5, Evaluation protocol=Zero-shot2026.02 | 33.19 | 33.02 | 44.36 | 68.23 | 41.15 | 48.86 | 51.32 | 35.59 | 30.14 | 36.54 | 42.24 | |
| PPLBackbone=GPT-2 Large, Optimizer=Muon, Training update tokens=30B, Data quality subset score=4+5, Evaluation protocol=Zero-shot2026.02 | 33.17 | 33.87 | 42.25 | 67.63 | 40.33 | 48.22 | 50.79 | 28.47 | 29.48 | 38.46 | 41.27 | |
| QuRatingBackbone=GPT-2 XL, Optimizer=Muon, Training update tokens=30B, Data quality subset score=4+5, Evaluation protocol=Zero-shot2026.02 | 33.1 | 33.58 | 44.22 | 66.7 | 39.97 | 49.64 | 50.09 | 32.54 | 28.99 | 36.54 | 41.54 | |
| FineWeb-EduBackbone=GPT-2 XL, Optimizer=Muon, Training update tokens=30B, Data quality subset score=4+5, Evaluation protocol=Zero-shot2026.02 | 32.94 | 33.64 | 43.14 | 68.28 | 39.61 | 51.3 | 52.73 | 32.2 | 31.37 | 36.54 | 42.18 | |
| RandomBackbone=GPT-2 Large, Optimizer=Muon, Training update tokens=30B, Data quality subset score=4+5, Evaluation protocol=Zero-shot2026.02 | 32.92 | 33.38 | 41.95 | 67.46 | 38.84 | 47.75 | 53.97 | 29.15 | 30.79 | 36.54 | 41.28 | |
| GREATSBackbone=GPT-2 Large, Optimizer=Muon, Training update tokens=30B, Data quality subset score=4+5, Evaluation protocol=Zero-shot2026.02 | 32.73 | 34.38 | 45.86 | 70.95 | 39.3 | 50.36 | 44.62 | 24.75 | 32.92 | 38.46 | 41.43 | |
| DCLM-FastTextBackbone=GPT-2 Large, Optimizer=Muon, Training update tokens=30B, Data quality subset score=4+5, Evaluation protocol=Zero-shot2026.02 | 32.64 | 33.67 | 41.66 | 66.38 | 38.74 | 51.3 | 49.38 | 30.85 | 31.04 | 36.54 | 41.22 | |
| OPUSBackbone=GPT-2 XL, Optimizer=Muon, Training update tokens=30B, Data quality subset score=3, Evaluation protocol=Zero-shot2026.02 | 32.62 | 33.11 | 50.54 | 72.2 | 41.04 | 51.46 | 47.62 | 30.85 | 35.63 | 54.81 | 44.99 | |
| UltraFinewebBackbone=GPT-2 Large, Optimizer=Muon, Training update tokens=30B, Data quality subset score=4+5, Evaluation protocol=Zero-shot2026.02 | 32.6 | 33.02 | 40.7 | 66.05 | 38.23 | 49.72 | 48.32 | 30.17 | 29.24 | 36.54 | 40.46 | |
| PPLBackbone=GPT-2 XL, Optimizer=Muon, Training update tokens=30B, Data quality subset score=4+5, Evaluation protocol=Zero-shot2026.02 | 32.6 | 33.58 | 46.14 | 69.1 | 40.33 | 51.7 | 50.79 | 30.17 | 31.78 | 36.54 | 42.27 | |
| DSIRBackbone=GPT-2 Large, Optimizer=Muon, Training update tokens=30B, Data quality subset score=4+5, Evaluation protocol=Zero-shot2026.02 | 32.54 | 33.54 | 41.07 | 67.95 | 39.36 | 47.28 | 48.68 | 33.9 | 29.57 | 38.46 | 41.24 | |
| OPUSBackbone=GPT-2 Large, Optimizer=Muon, Training update tokens=30B, Data quality subset score=4+5, Evaluation protocol=Zero-shot2026.02 | 32.17 | 33.38 | 42.52 | 67.3 | 39.51 | 51.07 | 54.14 | 30.85 | 31.04 | 38.46 | 42.04 | |
| FineWeb-EduBackbone=GPT-2 Large, Optimizer=Muon, Training update tokens=30B, Data quality subset score=4+5, Evaluation protocol=Zero-shot2026.02 | 32 | 33.46 | 39.95 | 64.74 | 39.87 | 50.51 | 52.2 | 29.15 | 30.3 | 36.54 | 40.87 | |
| RandomBackbone=GPT-2 XL, Optimizer=Muon, Training update tokens=30B, Data quality subset score=3, Evaluation protocol=Zero-shot2026.02 | 31.92 | 33.56 | 48.39 | 70.13 | 41.1 | 48.86 | 44.86 | 28.47 | 34.23 | 36.54 | 41.81 | |
| QuRatingBackbone=GPT-2 Large, Optimizer=Muon, Training update tokens=30B, Data quality subset score=4+5, Evaluation protocol=Zero-shot2026.02 | 31.32 | 34.07 | 41.7 | 66.92 | 39.71 | 47.83 | 50.79 | 32.88 | 31.94 | 36.54 | 41.37 | |
| RandomBackbone=GPT-2 Large, Optimizer=Muon, Training update tokens=30B, Data quality subset score=3, Evaluation protocol=Zero-shot2026.02 | 30.52 | 33.16 | 43.95 | 68.87 | 40.58 | 49.02 | 48.39 | 25.08 | 35.54 | 36.54 | 41.17 | |
| OPUSBackbone=GPT-2 Large, Optimizer=Muon, Training update tokens=30B, Data quality subset score=3, Evaluation protocol=Zero-shot2026.02 | 30.39 | 34.31 | 46.36 | 70.51 | 39.41 | 50.2 | 45.33 | 28.47 | 33.74 | 38.46 | 41.72 |