calibration.json
1007 B · 44 lines · json Raw
1 {
2 "version": 1,
3 "per_kind": {
4 "noul": 1.0022178236691426,
5 "choice": 0.9840233476014968,
6 "score": 1.012184473350984
7 },
8 "per_kind_family": {},
9 "fit": {
10 "noul": {
11 "n": 4363,
12 "T": 1.0022178236691426,
13 "kl_before": 0.004595278762280941,
14 "kl_after": 0.0045948573388159275
15 },
16 "choice": {
17 "n": 3418,
18 "T": 0.9840233476014968,
19 "kl_before": 0.04468083754181862,
20 "kl_after": 0.04462483152747154
21 },
22 "score": {
23 "n": 3173,
24 "T": 1.012184473350984,
25 "kl_before": 0.022512147203087807,
26 "kl_after": 0.022472143173217773
27 }
28 },
29 "diagnostic_calibration_split": {
30 "raw": {
31 "kl": 0.02229316346347332,
32 "ece": 0.0014972193181917403,
33 "mce": 0.0034518837928771973
34 },
35 "calibrated": {
36 "kl": 0.02226393297314644,
37 "ece": 0.0013502779446837457,
38 "mce": 0.005257606506347656
39 }
40 },
41 "source": "checkpoints/s2_9b_epoch1/best",
42 "fit_rows": 10954,
43 "d1_excluded": true
44 }