tokenfold-select-v2 / TRAINING-RECIPE.json
snchimata's picture
Release Tokenfold Select v2 ONNX and Safetensors preview
78736d9 verified
Raw History Blame Contribute Delete
2.66 kB
{
"seed": 0,
"epochs": 3,
"batch": 1,
"accumulation": 8,
"optimizer": "AdamW",
"learning_rate": 1e-05,
"weight_decay": 0.01,
"betas": [
0.9,
0.999
],
"eps": 1e-08,
"schedule": "constant",
"precision": "fp32",
"backend": "sdpa",
"gradient_checkpointing": true,
"clip_grad_norm": 1.0,
"student_token_cap": 8192,
"scoring_window_profile": "source-tail-1536-capacity-v2",
"scoring_context_token_cap": 1536,
"scoring_window_code_sha256": "fa879fb896e305a60ff10f3f5a6a23887da11b9741a210e2dc058154e148d929",
"supervision": "row-global optional class-balanced BCE plus mean window-local hardest optional pair softplus",
"feature_weight": 0.1,
"distillation": "original foundation encoder pooled features; no teacher head/logits",
"feature_reduction": "row-mean cosine distance over all original groups",
"gradient_reduction": "row mean then accumulation-block mean; one live window graph",
"checkpoint_criterion": "first minimum DEV balanced BCE plus window-local hard-pair row mean; provisional only",
"early_stopping": "none",
"test_access": false,
"qualification": false,
"objective_profile": "optional-balanced-window-hardpair-headwarm-v4",
"progress_publication": "immutable ready-marked snapshots plus terminal-only report",
"hard_pair_weight": 1.0,
"hard_pair_margin": 1.0,
"hard_pair_loss": "softplus(margin + max known optional negative - min known optional positive)",
"hard_pair_reduction": "row mean over windows containing both known optional classes; zero if none; window-local, not cross-window",
"hard_pair_role_admission": "at least one pair-bearing window in each TRAIN/DEV role; per-family coverage separately reviewed",
"head_warmup": {
"source": "fresh frozen ORIGINAL initializer encoder; TRAIN only; no saved cache or trained parent",
"objective": "row-global optional balanced BCE plus L2/2 on weight and bias",
"l2": 0.0001,
"device": "cpu",
"precision": "float64",
"optimizer": "LBFGS",
"learning_rate": 1.0,
"max_iter": 200,
"max_eval": 500,
"history_size": 20,
"line_search_fn": "strong_wolfe",
"tolerance_grad": 1e-07,
"tolerance_change": 1e-09,
"closure_cap": 512,
"max_design_bytes": 134217728,
"installation": "head-only FP32 copy; encoder unchanged; before teacher and joint AdamW creation",
"selection": "warm-only state ineligible; first minimum among three full joint DEV epochs",
"convergence_certified": false
},
"joint_gradient_evidence": "per AdamW update head/encoder preclip and total pre/postclip L2 norms"
}