JEV-9B / calibration.json
cloudyu's picture
autotrust/JEV: model weights, System 1 adapter + decision head, vLLM adapter, calibration and evaluation reports
c3ee780
Raw History Blame Contribute Delete
1.01 kB
{
"version": 1,
"per_kind": {
"noul": 1.0022178236691426,
"choice": 0.9840233476014968,
"score": 1.012184473350984
},
"per_kind_family": {},
"fit": {
"noul": {
"n": 4363,
"T": 1.0022178236691426,
"kl_before": 0.004595278762280941,
"kl_after": 0.0045948573388159275
},
"choice": {
"n": 3418,
"T": 0.9840233476014968,
"kl_before": 0.04468083754181862,
"kl_after": 0.04462483152747154
},
"score": {
"n": 3173,
"T": 1.012184473350984,
"kl_before": 0.022512147203087807,
"kl_after": 0.022472143173217773
}
},
"diagnostic_calibration_split": {
"raw": {
"kl": 0.02229316346347332,
"ece": 0.0014972193181917403,
"mce": 0.0034518837928771973
},
"calibrated": {
"kl": 0.02226393297314644,
"ece": 0.0013502779446837457,
"mce": 0.005257606506347656
}
},
"source": "checkpoints/s2_9b_epoch1/best",
"fit_rows": 10954,
"d1_excluded": true
}