Hugging Face's logo

sidbaines
/
pc-exp2-motivation-generalization

Safetensors
Model card Files
xet
 Community
sidbaines commited on
Commit
c910448
·
verified ·
1 Parent(s): 01ec63a

Upload folder using huggingface_hub

Browse files
Files changed (6) hide show
  1. fear_dpo/partb/fear_dpo_epoch1_decisiveness.json +59 -0
  2. fear_dpo/partb/fear_dpo_epoch1_efficacy.json +15 -0
  3. fear_dpo/partb/fear_dpo_epoch2_decisiveness.json +59 -0
  4. fear_dpo/partb/fear_dpo_epoch2_efficacy.json +15 -0
  5. fear_dpo/partb/fear_dpo_epoch3_decisiveness.json +59 -0
  6. fear_dpo/partb/fear_dpo_epoch3_efficacy.json +15 -0
fear_dpo/partb/fear_dpo_epoch1_decisiveness.json ADDED
@@ -0,0 +1,59 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "decisiveness": 0.8249730105158483,
3
+ "decisiveness_raw": 0.9925499246846877,
4
+ "unidim_fit_brier": 0.08137543938073905,
5
+ "unidim_fit_log_loss": 0.24769266265915005,
6
+ "transitivity_fas": 0.9112193589483801,
7
+ "transitivity_triad": 0.9415568602902584,
8
+ "order_consistency": 0.7599750859349228,
9
+ "q_agreement": 0.4999338698994261,
10
+ "q_agreement_absdiff": 0.7525879896175965,
11
+ "q_sign_agreement": 0.7552447552447552,
12
+ "mu_std_diagnostic": 4.989696090953074,
13
+ "_ci": {
14
+ "decisiveness": [
15
+ NaN,
16
+ NaN
17
+ ],
18
+ "decisiveness_raw": [
19
+ NaN,
20
+ NaN
21
+ ],
22
+ "unidim_fit_brier": [
23
+ NaN,
24
+ NaN
25
+ ],
26
+ "unidim_fit_log_loss": [
27
+ NaN,
28
+ NaN
29
+ ],
30
+ "transitivity_fas": [
31
+ NaN,
32
+ NaN
33
+ ],
34
+ "transitivity_triad": [
35
+ NaN,
36
+ NaN
37
+ ],
38
+ "order_consistency": [
39
+ NaN,
40
+ NaN
41
+ ],
42
+ "q_agreement": [
43
+ NaN,
44
+ NaN
45
+ ],
46
+ "q_agreement_absdiff": [
47
+ NaN,
48
+ NaN
49
+ ],
50
+ "q_sign_agreement": [
51
+ NaN,
52
+ NaN
53
+ ],
54
+ "mu_std_diagnostic": [
55
+ NaN,
56
+ NaN
57
+ ]
58
+ }
59
+ }
fear_dpo/partb/fear_dpo_epoch1_efficacy.json ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "arm": "fear_dpo_epoch1",
3
+ "rows": 24,
4
+ "summary": [
5
+ {
6
+ "arm": "fear_dpo_epoch1",
7
+ "dimension": "sycophancy",
8
+ "mean": 6.75,
9
+ "ci_lo": 6.0203,
10
+ "ci_hi": 7.4797,
11
+ "n": 24,
12
+ "n_unscored": 0
13
+ }
14
+ ]
15
+ }
fear_dpo/partb/fear_dpo_epoch2_decisiveness.json ADDED
@@ -0,0 +1,59 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "decisiveness": 0.8169865422077078,
3
+ "decisiveness_raw": 0.983767750326317,
4
+ "unidim_fit_brier": 0.08849969334434242,
5
+ "unidim_fit_log_loss": 0.26873436389681055,
6
+ "transitivity_fas": 0.9102443592239899,
7
+ "transitivity_triad": 0.9422620068180338,
8
+ "order_consistency": 0.7621709895889291,
9
+ "q_agreement": 0.4495462924435902,
10
+ "q_agreement_absdiff": 0.7171302650559261,
11
+ "q_sign_agreement": 0.7039473684210527,
12
+ "mu_std_diagnostic": 4.862116103944152,
13
+ "_ci": {
14
+ "decisiveness": [
15
+ NaN,
16
+ NaN
17
+ ],
18
+ "decisiveness_raw": [
19
+ NaN,
20
+ NaN
21
+ ],
22
+ "unidim_fit_brier": [
23
+ NaN,
24
+ NaN
25
+ ],
26
+ "unidim_fit_log_loss": [
27
+ NaN,
28
+ NaN
29
+ ],
30
+ "transitivity_fas": [
31
+ NaN,
32
+ NaN
33
+ ],
34
+ "transitivity_triad": [
35
+ NaN,
36
+ NaN
37
+ ],
38
+ "order_consistency": [
39
+ NaN,
40
+ NaN
41
+ ],
42
+ "q_agreement": [
43
+ NaN,
44
+ NaN
45
+ ],
46
+ "q_agreement_absdiff": [
47
+ NaN,
48
+ NaN
49
+ ],
50
+ "q_sign_agreement": [
51
+ NaN,
52
+ NaN
53
+ ],
54
+ "mu_std_diagnostic": [
55
+ NaN,
56
+ NaN
57
+ ]
58
+ }
59
+ }
fear_dpo/partb/fear_dpo_epoch2_efficacy.json ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "arm": "fear_dpo_epoch2",
3
+ "rows": 24,
4
+ "summary": [
5
+ {
6
+ "arm": "fear_dpo_epoch2",
7
+ "dimension": "sycophancy",
8
+ "mean": 7.125,
9
+ "ci_lo": 6.1218,
10
+ "ci_hi": 8.1282,
11
+ "n": 24,
12
+ "n_unscored": 0
13
+ }
14
+ ]
15
+ }
fear_dpo/partb/fear_dpo_epoch3_decisiveness.json ADDED
@@ -0,0 +1,59 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "decisiveness": 0.8291821802568966,
3
+ "decisiveness_raw": 0.9775292605760791,
4
+ "unidim_fit_brier": 0.07180464767475643,
5
+ "unidim_fit_log_loss": 0.24387921484557265,
6
+ "transitivity_fas": 0.917065694669698,
7
+ "transitivity_triad": 0.9258820618879724,
8
+ "order_consistency": 0.7458697571783026,
9
+ "q_agreement": 0.5116229122276893,
10
+ "q_agreement_absdiff": 0.7563423738499957,
11
+ "q_sign_agreement": 0.7534246575342466,
12
+ "mu_std_diagnostic": 4.7123481084918355,
13
+ "_ci": {
14
+ "decisiveness": [
15
+ NaN,
16
+ NaN
17
+ ],
18
+ "decisiveness_raw": [
19
+ NaN,
20
+ NaN
21
+ ],
22
+ "unidim_fit_brier": [
23
+ NaN,
24
+ NaN
25
+ ],
26
+ "unidim_fit_log_loss": [
27
+ NaN,
28
+ NaN
29
+ ],
30
+ "transitivity_fas": [
31
+ NaN,
32
+ NaN
33
+ ],
34
+ "transitivity_triad": [
35
+ NaN,
36
+ NaN
37
+ ],
38
+ "order_consistency": [
39
+ NaN,
40
+ NaN
41
+ ],
42
+ "q_agreement": [
43
+ NaN,
44
+ NaN
45
+ ],
46
+ "q_agreement_absdiff": [
47
+ NaN,
48
+ NaN
49
+ ],
50
+ "q_sign_agreement": [
51
+ NaN,
52
+ NaN
53
+ ],
54
+ "mu_std_diagnostic": [
55
+ NaN,
56
+ NaN
57
+ ]
58
+ }
59
+ }
fear_dpo/partb/fear_dpo_epoch3_efficacy.json ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "arm": "fear_dpo_epoch3",
3
+ "rows": 24,
4
+ "summary": [
5
+ {
6
+ "arm": "fear_dpo_epoch3",
7
+ "dimension": "sycophancy",
8
+ "mean": 6.9583,
9
+ "ci_lo": 5.9609,
10
+ "ci_hi": 7.9558,
11
+ "n": 24,
12
+ "n_unscored": 0
13
+ }
14
+ ]
15
+ }