1M-param 4-digit addition model + full results (phase diagram, extrapolation, bake-off)
4a6c57b verified | step,loss,lr,eval_acc,elapsed_s,tok_per_s | |
| 610,1.0380,5.96e-04,0.0000,15.3,326115 | |
| 1220,0.6602,5.67e-04,0.6100,30.3,330209 | |
| 1830,0.6154,5.13e-04,0.9060,45.1,332162 | |
| 2440,0.6263,4.39e-04,0.9760,60.1,332463 | |
| 3050,0.6127,3.53e-04,0.9740,75.2,332049 | |
| 3660,0.6131,2.64e-04,1.0000,90.4,331842 | |
| 4270,0.6106,1.83e-04,0.9980,105.4,331818 | |
| 4880,0.6108,1.17e-04,1.0000,120.4,331979 | |
| 5490,0.6092,7.48e-05,1.0000,135.5,332033 | |
| 6100,0.6057,6.00e-05,1.0000,150.5,332027 | |
| 6103,0.6028,6.00e-05,1.0000,151.8,329437 | |