| ======================================================= |
| EDDIE TRAINING STATUS — 2x T4 DDP (SPEED EXP) |
| ======================================================= |
| Updated : 2026-09-07 20:06:43 UTC |
| Epoch : 9060 |
| Train CE : 0.809283 |
| Best CE : 0.801314 |
| Cortex base : 0.799489 |
| U-loss (Huber): 0.110507 |
| Total loss : 0.919790 |
| ------------------------------------------------------- |
| u target : excess (signed excess of p; ~0 = trust p) |
| excess mean : +0.0045 (p calibrated on avg if ~0) |
| u mean : -0.0403 (should track excess mean) |
| u corr : 0.0033 (Pearson u vs realized excess) |
| u R^2 : -0.0312 (>0 = beats constant) |
| u spread : -0.0035 (excess top-q u minus bottom-q u; want >0) |
| const Huber : 0.1143 (u useful if U-loss < this: YES) |
| ------------------------------------------------------- |
| Speed : 373 samples/s | 1.4 min/epoch |
| Batch/GPU : 32 | GPUs: 2 |
| Grad Accum : 8 | Eff BS: 512 |
| Samples/ep : 32,000 |
| ------------------------------------------------------- |
| Architecture : CortexEDL |
| Params : 3,058,052 |
| Context : 1024 bars | Features: 5 |
| Total windows: 501,051,135 |
| Note: no held-out validation split -- every training example |
| is drawn without replacement from ~2.4B windows, so batches |
| are effectively unseen data by construction. |
| ======================================================= |
|
|
