Evidence Based Settings Based on Target Text Complexity

These recommendations are data-driven recommendations based on the analysis described under Study: Randomness + TFS --> Readability.

Basically these are based on our test runs with one single prompt, modifying only TFS + randomness settings and then rating readability (data collection still ongoing!). The data was then analyzed via (polynomial) regression.

Take care though: We make no guarantees this works as intended. The effect of these settings might change considerably, depending on what is in your context. Also, as indicated by Basileus, we do not know yet how these settings influence story quality (coherence, creativity, etc.). And of course, the "predicted readability" predicts only the average predictability over many outputs.

All that being said, feel free to try out these settings! And if you do you are more than welcome to share your insights in #community-research .

Find the recommendations here.

Different generator settings

Also check #generation-settings in the NAI discord, I currently have no time to also summarize stuff in there. (Note that NAI let's you easily import/export generation settings. I only write them down here so we can begin to see patterns).

Also note that AWK's setting discord link (new candidate for default) seems to look promising.

Monky's preferred settings

(currently slightly prone to being repetitive after a few paragraphs of AI output)

Variable	Setting
Randomness	0.8
Max Output Length	60
Min Output Length	20
Top-K Sampling	disabled
Nucleus	disabled
Tail-Free Sampling	0.5
Repition Penalty	1.2
Repition Penalty Range	512
Repitition Penalty Slope	4.05

Shincore's settings

(alteration of Monkys)

Variable	Setting
Randomness	0.8
Max Output Length	60
Min Output Length	30
Top-K Sampling	disabled
Nucleus	disabled
Tail-Free Sampling	0.8
Repition Penalty	1.3
Repition Penalty Range	512
Repitition Penalty Slope	2.97

Vegetables are not geese settings

Link

Variable	Setting
Randomness	0.45
Max Output Length	40
Min Output Length	20
Top-K Sampling	disabled
Nucleus	disabled
Tail-Free Sampling	0.992
Repition Penalty	3
Repition Penalty Range	1024
Repitition Penalty Slope	3.6

I tested multiple runs of the same thing at randomness .45 and TFS at .985 .990 .993 .995 .997 .999 side by side

and I tested all the values and found that the sweetspot for the style of generation I wanted was more on the .990 side than the .999 side, but higher than .985
and tested more and started taking .001 off and stuff, found a specific place where .993 made a logic error that .992 didn't and settled at .992