image
image

I made WinterGoliath and it felt a bit off compared to regular Goliath. I was wondering if left bias was really present in the models after that so I made an automatic benchmark using politiscales test.

Interpreting the data

  • b0/b1 : Internationalism/Nationalism
  • c0/c1 : Constructivism/Essentialism
  • e0/e1 : Ecology/Production
  • j0/j1 : Rehabilitative Justice/Punitive Justice
  • m0/m1 : Regulation/Laissez-faire
  • p0/p1 : Communism/Capitalism
  • s0/s1 : Progressive/Conservative
  • t0/t1 : Revolution/Reform
  • reli : Religiousness
  • comp : Belief in a worldwide conspiracy
  • prag : Pragmatism
  • mona : Monarchism
  • vega : VEGAN
  • anar : Anarchism
  • femi: : Radical feminism

Suggestions for calculations

  • Whackiness of the model=anar+comp+mona+reli
  • Certainty=(x0+x1), calculate average of all values
  • Bias towards value(logratio)=LOG(x0/x1;2)
  • Left-right bias=b_logratio+c_logratio+e_logratio+j_logratio+m_logratio+p_logratio+s_logratio-t_logratio

Results

politiscales-llm.png

  • Mixtral-instruct0.1 hates violence(-3.1 logratio_t!!!), religion(0.4%), pragmatic politics(1.4%), monarchy(0.6%), anarchy(5.7%) and conspiracy theories(0.2%); likes regulation(1.5 logratio_t), globalism(1.4) and ecology(0.64). The MOST vegan(72%!!!) model tested so far. EXTREME left-wing bias.
  • LLAMA2-70b is fairly neutral, with slight bias to the left.
  • Dicephal-123(self-merge of llama2-70B) is also fairly neutral, with slight bias to the right. Sadly, it is the most right-wing model that I tested.
  • Xwin, Goliath, Nous-Hermes have low left-wing bias.
  • DoubleGold and WinterGoliath have medium left-wing bias.
  • WinterGoddess and Euryale have high left-wing bias.

My suspicions were correct, WinterGoliath has stronger left wing bias than Goliath and is less whacky, I cannot consider it an upgrade, just a sidegrade.

Limitations

The tests were done without any special prompting. How models perform on the test does not reflect how they perform with the right prompting, it just shows the bias of the model.

This test does NOT measure censorship.

Downloads last month
0
Edit dataset card