Question Answering on SciQ (Acc)
89.1AccuracyParallax
Evaluation Results
| Method | Links | |
|---|---|---|
| ParallaxSize=1.7B, Optimizer=Muon, RoPE on ρ=✗2026.05 | 89.1 | |
| ParallaxSize=1.7B, Optimizer=Muon, RoPE on ρ=✓2026.05 | 88.7 | |
| TransformerSize=1.7B, Optimizer=Muon, RoPE on ρ=—2026.05 | 88 | |
| ParallaxSize=0.6B, Optimizer=Muon, RoPE on ρ=✗2026.05 | 83.8 | |
| ParallaxSize=0.6B, Optimizer=Muon, RoPE on ρ=✓2026.05 | 83.6 | |
| Parallax†Size=0.6B, Optimizer=Muon, RoPE on ρ=✓2026.05 | 82.4 | |
| Transformer†Size=0.6B, Optimizer=Muon, RoPE on ρ=—2026.05 | 82.3 | |
| TransformerSize=0.6B, Optimizer=Muon, RoPE on ρ=—2026.05 | 81.2 | |
| TransformerSize=0.6B, Optimizer=Cosine AdamW, RoPE on ρ=—2026.05 | 80.5 | |
| TransformerSize=0.6B, Optimizer=WSD AdamW, RoPE on ρ=—2026.05 | 80.5 | |
| ParallaxSize=0.6B, Optimizer=WSD AdamW, RoPE on ρ=✓2026.05 | 78.8 | |
| Kimi DeltaAttnSize=0.6B, Optimizer=Muon, RoPE on ρ=—2026.05 | 78.5 | |
| Gated DeltaNetSize=0.6B, Optimizer=Muon, RoPE on ρ=—2026.05 | 78.5 | |
| ParallaxSize=0.6B, Optimizer=Cosine AdamW, RoPE on ρ=✓2026.05 | 74.4 |