Natural Language Understanding on HellaSwag
85.6AccuracyNLS
Evaluation Results
| Method | Links | |
|---|---|---|
| NLSModel=Llama 3 70B, Precision=W8A8, Time Steps (T)=22026.05 | 85.6 | |
| PrefixQModel=Llama 2 7B, Precision=W8A8, Time Steps (T)=-2026.05 | 76.44 | |
| PrefixQModel=Llama 2 7B, Precision=W6A6, Time Steps (T)=-2026.05 | 74.06 | |
| Fixed (p = 4) + SPSequence Reduction=4.0, FLOPs/byte Reduction=2.12026.05 | 59.7 | |
| Entropy-based + SPSequence Reduction=6.3†, FLOPs/byte Reduction=2.72026.05 | 59.7 | |
| H-Net + SPSequence Reduction=5.3†, FLOPs/byte Reduction=2.42026.05 | 59.6 | |
| SpaceByte + SPSequence Reduction=6.3, FLOPs/byte Reduction=2.72026.05 | 59.5 | |
| Fixed (p = 8) + SPSequence Reduction=8.0, FLOPs/byte Reduction=2.62026.05 | 59.2 | |
| H-NetSequence Reduction=4.8, FLOPs/byte Reduction=3.42026.05 | 58.5 | |
| Tokenizer-basedSequence Reduction=3.7, FLOPs/byte Reduction=4.12026.05 | 58.4 | |
| Fixed (p = 16) + SPSequence Reduction=16.0, FLOPs/byte Reduction=2.92026.05 | 58 | |
| Fixed (p = 4)Sequence Reduction=4.0, FLOPs/byte Reduction=3.12026.05 | 57.9 | |
| Entropy-basedSequence Reduction=6.4, FLOPs/byte Reduction=4.02026.05 | 57.7 | |
| SpaceByteSequence Reduction=6.3, FLOPs/byte Reduction=4.02026.05 | 57.3 | |
| Byte-levelSequence Reduction=1.0, FLOPs/byte Reduction=1.02026.05 | 55.1 | |
| Fixed (p = 8)Sequence Reduction=8.0, FLOPs/byte Reduction=4.52026.05 | 54.5 | |
| SYNPROModel Size=1.1B, Alpha (organic token ratio)=10%2026.05 | 49.17 | |
| Fixed (p = 16)Sequence Reduction=16.0, FLOPs/byte Reduction=5.72026.05 | 47.3 | |
| UniqueModel Size=400M, Alpha (organic token ratio)=10%2026.05 | 43.03 | |
| FTProjection Type=Out-Proj, Number of Parameters=28.31M2024.11 | 40.89 | |
| FTProjection Type=In-Proj, Number of Parameters=56.62M2024.11 | 40.36 | |
| FTProjection Type=Both-Proj, Number of Parameters=84.94M2024.11 | 40.18 | |
| ProDiaLProjection Type=Both-Proj, Number of Parameters=2.42M2024.11 | 38.92 | |
| SYNPROModel Size=400M, Alpha (organic token ratio)=10%2026.05 | 38.68 | |
| StrongProjection Type=Baseline, Number of Parameters=3.80M2024.11 | 38.66 | |
| LoRAProjection Type=In-Proj, Number of Parameters=1.48M2024.11 | 38.46 | |
| ProDiaLProjection Type=In-Proj, Number of Parameters=1.49M2024.11 | 38.41 | |
| LoRAProjection Type=Both-Proj, Number of Parameters=2.36M2024.11 | 38.33 | |
| Full-FTProjection Type=Baseline, Number of Parameters=130.00M2024.11 | 38.23 | |
| ProDiaLProjection Type=Out-Proj, Number of Parameters=0.92M2024.11 | 38.19 | |
| DoRAProjection Type=Both-Proj, Number of Parameters=2.45M2024.11 | 38.13 | |
| DoRAProjection Type=In-Proj, Number of Parameters=1.55M2024.11 | 38.08 | |
| LoRAProjection Type=Out-Proj, Number of Parameters=0.89M2024.11 | 37.3 | |
| DoRAProjection Type=Out-Proj, Number of Parameters=0.90M2024.11 | 37.19 | |
| BitFitProjection Type=Baseline, Number of Parameters=0.07M2024.11 | 35.69 |