L3.3-Electra-R1-70b-Elarablated-v0.1

This model has been "Elarablated"; that is, I've used a special kind of training to specifically target and remove certain railroaded tokens (cliches, slop, call them what you will). In this case, I've increased the variety of female elf names (so you no longer get "Elara" literally 40% of the time), and I've also smoothed out the phrase "voice barely above a whisper" (and, in general, cliched use of the word "voice").

Here are some screens showing token probabilities:

Before Elarablation (note how the token probabilities railroad straight down "barely above a whisper"):

After Elarablation (note the significantly more even token probabilties):

This is still in a very early testing phase. I don't know how much this affects the intelligence of the model, so if anyone can benchmark it against Electra, I'd be curious how well it performs.

For the Elarablation code, see my github repo, here:

https://github.com/envy-ai/elarablate