| # PSEng-14B-preview: iteration 500 selected | |
| Selected on the validation curve alone, before any dev score existed. | |
| | iter | val loss | | |
| |---|---| | |
| | 150 | 0.275 | | |
| | 200 | 0.274 | | |
| | 250 | 0.274 | | |
| | 300 | 0.259 | | |
| | 350 | 0.270 | | |
| | 400 | 0.292 | | |
| | 450 | 0.253 | | |
| | **500** | **0.243** | | |
| | 550 | 0.286 | | |
| Rule, as in v1: minimum validation loss; ties break toward fewer iterations. | |
| Iteration 500 is the minimum over the observed curve. | |
| Training was stopped at iteration ~560 of a planned 1,200 rather than run to | |
| completion. Reason, recorded plainly: sustained 14B QLoRA on this machine | |
| drains the battery ~30 points/hour even on AC (the 96W adapter cannot keep | |
| up), and the curve had shown no improvement past 500 across three validation | |
| points. The remaining iterations were judged not worth the power. This is a | |
| preview artifact; the Phase 2 sweep revisits schedule length on a rested | |
| machine. | |
| Note the curve is noisy (0.259 at 300, 0.292 at 400, 0.243 at 500), so "500 is | |
| best" is a weak claim over a shallow, jittery minimum, not evidence of a | |
| well-converged optimum. | |
| Base: models/phi-4-4bit (QLoRA; adapters trained against the quantized base). | |
| Corpus: data/phi_lite, 517 train / 43 valid. | |