| optimizer,category,lr,final_val_loss,final_val_ppl,final_bleu,best_bleu,best_val_loss_at_frac,train_time_min,opt_state_MB,state_per_param_bytes,speedup_vs_adamw |
| adamw,AdamW (baseline),0.001,3.8300798034667967,46.066214331135704,4.3420847588880305,4.3420847588880305,0.9999808677129902,11.125547536214192,209.05078125,8.0, |
| mars,Variance-reduced AdamW variant,0.002,3.7394594593048094,42.07524063720369,4.215608656483604,4.215608656483604,0.9999808677129902,12.17287700176239,313.55078125,11.999028346133004,1.2605100274902472 |
| lion,Memory-efficient,0.0002,4.102442771911621,60.487865333032985,3.3781852579367238,3.3781852579367238,0.9999808677129902,11.420517086982727,104.525390625,4.0, |
| muon,Matrix-based,0.02,3.545899326324463,34.67085173543859,3.997071143178155,4.485001421756144,0.9999808677129902,15.120738577842712,137.05078125,5.244688603621279,1.816714649962729 |
| sophia,Hessian-approximation (bonus),0.0002,4.095900161743164,60.09340860561834,3.7665308993611517,3.7665308993611517,0.9999808677129902,12.026670670509338,209.05078125,8.0, |
|
|
