Phase 4: v1 training run results (val loss 1.52, ppl ~4.6)

- reports/loss-v1.csv: full train/val loss log over 20k steps - reports/loss-v1.txt: ASCII loss curve (steep descent then flattening tail) Training ran ~58 min on M1 MPS at 5.7 it/s. No overfitting (train~=val).
2026-07-12 12:35:47 -04:00
parent 72595e274b
commit ba73c8a7e8
2 changed files with 62 additions and 0 deletions
@@ -0,0 +1,40 @@
+iter,train_loss,val_loss,lr
+500,2.8285,2.8171,0.000600
+1000,2.3311,2.3267,0.000598
+1500,2.1135,2.1026,0.000594
+2000,2.0013,2.0002,0.000589
+2500,1.9253,1.9230,0.000582
+3000,1.8811,1.8752,0.000574
+3500,1.8399,1.8429,0.000564
+4000,1.8119,1.8128,0.000552
+4500,1.7775,1.7905,0.000540
+5000,1.7589,1.7616,0.000525
+5500,1.7466,1.7440,0.000510
+6000,1.7248,1.7261,0.000494
+6500,1.7038,1.7093,0.000476
+7000,1.6980,1.7023,0.000458
+7500,1.6829,1.6890,0.000438
+8000,1.6747,1.6796,0.000418
+8500,1.6572,1.6673,0.000398
+9000,1.6506,1.6452,0.000377
+9500,1.6408,1.6319,0.000356
+10000,1.6290,1.6307,0.000334
+10500,1.6202,1.6215,0.000313
+11000,1.6126,1.6166,0.000292
+11500,1.6132,1.6014,0.000271
+12000,1.5962,1.5903,0.000250
+12500,1.5851,1.5741,0.000230
+13000,1.5807,1.5864,0.000210
+13500,1.5666,1.5754,0.000191
+14000,1.5693,1.5663,0.000173
+14500,1.5564,1.5673,0.000156
+15000,1.5571,1.5529,0.000141
+15500,1.5415,1.5444,0.000126
+16000,1.5441,1.5388,0.000113
+16500,1.5396,1.5356,0.000101
+17000,1.5360,1.5270,0.000090
+17500,1.5305,1.5278,0.000081
+18000,1.5281,1.5263,0.000073
+18500,1.5192,1.5244,0.000068
+19000,1.5191,1.5201,0.000063
+19500,1.5122,1.5198,0.000061
@@ -0,0 +1,22 @@
+loss curve — v1  (final train 1.5122, val 1.5198 @ iter 19500)
+ 2.829 |*                                                               
+ 2.751 |                                                                
+ 2.674 |                                                                
+ 2.596 |                                                                
+ 2.519 |                                                                
+ 2.441 |                                                                
+ 2.364 |  *                                                             
+ 2.286 |                                                                
+ 2.209 |                                                                
+ 2.132 |   *                                                            
+ 2.054 |                                                                
+ 1.977 |     *                                                          
+ 1.899 |       **                                                       
+ 1.822 |          * **                                                  
+ 1.744 |             . * ** *                                           
+ 1.667 |                    . ** * ** * *                               
+ 1.590 |                                 * ** * ** * ** *               
+ 1.512 |                                                  ** * ** * ** *
+       +----------------------------------------------------------------
+        500                                                        19500
+  legend: '.' train   '*' val   (x: iter, y: loss)