Last-word prediction on LAMBADA (en, de, es, fr, it average)
18.2Zero-shot AccuracyMGPT-2
Evaluation Results
| Method | Links | |
|---|---|---|
| MGPT-2#Params=124M, Patch size=-, zero-shot evaluation=true2026.04 | 18.2 | |
| MIXAR stage2#Params=477M, Patch size=32×32, zero-shot evaluation=true2026.04 | 9.6 | |
| GPT-2#Params=124M, Patch size=-, zero-shot evaluation=true2026.04 | 8.3 | |
| PIXAR stage2#Params=113M, Patch size=8×8, zero-shot evaluation=true2026.04 | 6.3 | |
| MIXAR stage2#Params=116M, Patch size=32×32, zero-shot evaluation=true2026.04 | 6.3 | |
| MIXAR stage1#Params=477M, Patch size=32×32, zero-shot evaluation=true2026.04 | 2.9 | |
| MIXAR stage1#Params=116M, Patch size=32×32, zero-shot evaluation=true2026.04 | 1.2 |