| { | |
| "epochs": 3, | |
| "initialization": "Official jaredpalmer/kev-4b@485ace8703592fcf405488b262449990824cfed1, then SecJev-Corpus", | |
| "optimizer": "fresh AdamW, no loaded optimizer moments; retained continuously through all epochs", | |
| "global_question_batches": { | |
| "1": 128, | |
| "2": 192, | |
| "3": 192 | |
| }, | |
| "steps_per_epoch": [ | |
| 929, | |
| 619, | |
| 619 | |
| ], | |
| "lr": "epoch1 peak2e-5 with5%warmup/cosine to2e-6; epochs2+3 peak1e-5 with5%warmup from2e-6/cosine to1e-6", | |
| "objective": "per-scene mean cross entropy", | |
| "weight_decay": 0.01, | |
| "grad_clip": 1.0, | |
| "world_size": 9, | |
| "comparison": "same security recipe as 0.8B/2B, official Kev-4B warm start, nine GPUs with memory-bounded microbatches", | |
| "train_questions_per_epoch": 118818, | |
| "replay": 0, | |
| "augmentation": false, | |
| "no_truncation": true, | |
| "context_limit": 8192, | |
| "test_use": "all3finalepochs predeclared for audit; selection uses full development only" | |
| } | |