These recommendations are data-driven recommendations based on the analysis described here (discord link).
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.
The readability grade "measures of how easy your text is to read. Your readability score shows what grade level of students could understand and engage with your writing. For instance, a score of 7 means that seventh-grade students could read your work". Below you will find a short explanation of what certain readability grades mean and what setting you can try to get a text with similar complexity from NAI (these settings are for Sigurd v2!).
However, take care! I can make no guarantees these settings actually work as intended. Our data so far has tested these settings only on one prompt. The effect of these settings might change considerably, depending on what is in your context. We do not know yet how the relationship between sampling & randomness settings change, depending on input!
Also, these settings are likely to influence many other things besides readability ratings. We are just starting out with researching text complexity, because that is easy to rate. 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. As always, there is a lot of random variability with any setting. So the findings described below describe the general tendency of the output with these settings.
All that being said, feel free to try out these settings for yourself! And if you do you are more than welcome to share your insights in the community research channel. Reminder: These settings are for Sigurd v2!
"Punchy" (4-5)
(Ernest Hemingway, Cormac McCarthy, Susan Collins): concise, punchy language with straightforward sentence structure and word choice
TFS-Sampling actually seems to usually get you something more elaborate than this (all predicted values are > 5).
Try: TF-Sampling = 0.9, Randomness = 0.45 (predicted readability = 5.2)
For Nucleus Sampling = 0.9 try: randomness = 0.6 (predicted readability = 4.9)
If you prefer TFS take something with randomness values around 0.3 - 0.5 and TFS ranges of 0.6 - 0.9 (predicted readability = 5.2 - 54)
"Casual" (6)
(Hunter S. Thompson, Steven King, Stephanie Meyer): flows well, easy to read large quantities of without getting tired
Try: TF-Sampling = 0.65, Randomness = 0.5 (predicted readability = 6.0)
For Nucleus Sampling = 0.9 try: Randomness = 0.8 (predicted readability = 5.8)
With TFS-Sampling it depends on your preferred randomness setting: With randomness < 0.3, keep TFS on 0.6 or higher. For randomness values between 0.3-0.5, a TFS setting of 0.4 should get you casual complexity. If your randomness is 0.6-0.7, TFS settings of 0.4 or higher should all get you casual complexity.
If you want Nucleus Sampling = 0.9, keep your randomness between 0.75-0.85
"Poetic" (7)
(J.R.R. Tolkien, Dan Brown, Thomas Pynchon): enough complexity to have clever turns of phrase, without having to stop and think/reread
Try: TF-Sampling = 0.75, Randomness = 0.6 (predicted readability = 7.0)
For Nucleus Sampling = 0.9 try: Randomness = 0.95 (predicted readability = 7.3)
With TFS-Sampling you either want really low randomness (0.1-0.2) with TFS of 0.4, or randomness of 0.75-0.8 (should work on any TFS above 0.4)
If you want Nucleus Sampling = 0.9, keep your randomness between 0.9-0.95
"Complex" (8-9)
(James Patterson, F. Scott Fitzgerald, Leo Tolstoy): relatively complex language, variety in word choice; can sometimes be abstruse
Try: TF-Sampling = 0.8, Randomness = 0.9 (predicted readability = 8.7)
For Nucleus Sampling = 0.9 try: Randomness = 1.0 (predicted readability = 7.9)
You need high randomness for this (0.85-0.9) - TFS can be anything up from 0.4