LemonDecider / sft_receipt.json
niqun123123's picture
Sanitize metadata: remove internal experiment paths and codenames
c1bfc95 verified
Raw History Blame Contribute Delete
651 Bytes
{
"step": 24,
"dev_accuracy_t1": 0.7735191583633423,
"best_step": 3,
"best_dev_accuracy_t1": 0.7804877758026123,
"lr": 2.2827515498423204e-07,
"examples_seen": 6144,
"temperature_fit": {
"method": "NLL scan on dev 574 under fork _known() semantics, T grid [0.30,5.00]x0.05 (from cached T=1 probs)",
"best_T": 0.6,
"dev_nll_at_T": 0.56265,
"dev_brier_at_T": 0.28261,
"dev_nll_at_1": 0.6016,
"dev_brier_at_1": 0.29885,
"bound": false,
"reason": "best_T=0.60 != serving T=1.0; NOT applied — public-set re-eval required before adoption (dev acc unchanged at T*, argmax-invariant except noul boundary)"
}
}