Fuminides's picture
Submission data
54adb0b verified
|
Raw History Blame Contribute Delete
2.89 kB

Steering evaluation: CohereLabs/tiny-aya-fire (safetune.CAAModel)

alpha in units of the trait's high-low persona gap at the layer. Says: questionnaire score of the steered trait (1-5, no persona). Does: behaviour index of the trait in the direct dilemmas (0-100). mass: probability on the digit tokens (low = the model stops answering).

Baseline digit mass: questionnaire 0.98, dilemmas 0.93.

trait layer alpha says -a says 0 says +a says change does -a does 0 does +a does change mass q -/+ mass d -/+
O 16 4 3.00 3.27 3.10 +0.11 45 59 70 +25 0.98 / 0.98 0.92 / 0.95
C 24 8 3.22 3.30 3.45 +0.23 52 49 47 -5 0.96 / 0.95 0.88 / 0.95
E 26 4 3.33 3.44 3.41 +0.08 26 26 26 -0 0.97 / 0.97 0.94 / 0.93
A 24 4 3.12 3.14 3.13 +0.01 59 59 59 -0 0.94 / 1.00 0.91 / 0.95
N 16 8 3.17 3.14 3.15 -0.01 64 73 44 -20 0.85 / 0.73 0.80 / 0.43
H 20 8 2.96 3.12 3.26 +0.30 62 73 80 +17 0.95 / 0.94 0.82 / 0.96
Mach 35 4 3.28 3.35 3.35 +0.07 23 27 28 +5 0.89 / 0.90 0.83 / 0.62
Narc 24 8 3.29 3.32 3.47 +0.18 36 41 45 +9 1.00 / 0.91 0.97 / 0.91
Psych 24 8 2.85 2.81 2.85 +0.01 39 40 43 +4 0.86 / 0.98 0.90 / 0.96

Questionnaire grid: change of the steered trait's own score between -alpha and +alpha (digit mass at -/+).

trait Lbest a=4 Lbest a=8 L16 a=4 L16 a=8 L24 a=4 L24 a=8
O (best L31) +0.00 (0.98/0.95) -0.01 (0.97/0.85) +0.11 (0.98/0.98) +0.00 (0.91/0.99) +0.02 (0.98/0.96) +0.03 (0.96/0.91)
C (best L25) +0.10 (0.98/0.97) +0.22 (0.96/0.96) +0.07 (0.99/0.98) -0.01 (0.92/0.97) +0.14 (0.99/0.97) +0.23 (0.96/0.95)
E (best L26) +0.08 (0.97/0.97) -0.08 (0.89/0.39) +0.06 (0.99/0.88) +0.15 (0.34/0.83) +0.01 (0.99/0.95) -0.01 (0.97/0.39)
A (best L35) -0.00 (0.98/0.95) +0.01 (0.21/0.71) -0.11 (0.98/0.95) +0.10 (0.26/0.86) +0.01 (0.94/1.00) -0.03 (0.74/1.00)
N (best L35) -0.09 (0.99/0.95) -0.09 (0.96/0.76) -0.05 (0.98/0.98) -0.01 (0.85/0.73) -0.09 (0.99/0.96) -0.20 (1.00/0.85)
H (best L20) +0.13 (0.97/0.98) +0.30 (0.95/0.94) +0.01 (0.97/0.99) -0.06 (0.90/0.98) +0.03 (0.98/0.99) +0.01 (0.96/0.99)
Mach (best L35) +0.07 (0.89/0.90) +0.04 (0.10/0.02) +0.32 (0.41/0.62) -0.01 (0.00/0.01) +0.02 (0.95/0.89) -0.18 (0.03/0.54)
Narc (best L24) +0.07 (1.00/0.96) +0.18 (1.00/0.91) -0.18 (0.99/0.97) -0.11 (0.97/0.94) +0.07 (1.00/0.96) +0.18 (1.00/0.91)
Psych (best L35) -0.02 (0.95/0.98) -0.03 (0.86/0.79) -0.28 (0.97/0.97) -0.31 (0.88/0.59) -0.01 (0.96/0.99) +0.01 (0.86/0.98)

9552s