Some feedback on a great model

#1
by Zeldazackman - opened

I thought I'd toss in some quick and simple feedback.

This model is great, I mean really great.

It responds to inputs quite well and expresses prose and fluidity that's been lacking with other models that I've ran into (Mostly supernova and qwen 2.5 merges).

On-top of that, chunky lotus doesn't seem to have any issues when facing complex thought processes (EG. how many iterations of 'jane and joe each think of two different people's names, of who each of them think of two other people as well' as it can go down that process without tripping up the names for a bit) or following instructions like other models that I've ran into.

The model also handles dynamic temperatures really well and I haven't had to toggle mirostat at all for unique, well thought out, and generally creative responses.

It can go a little off the rails at temps above 1.6 in my testing (Nearly every model does to be quite frank) but I've found that having a smoothing curve of 2 with a smoothing factor of 0.35 keeps it more in line along with min p being tweaked down to 0.05.

These are the settings that I've been using and work well enough for my uses.

lHrWMSX.png

Thanks so much for the feedback! I'm really glad to hear you're enjoying the model, and I appreciate you sharing the settings that have worked well for you. I'll have to try those out for myself!

So I've been using your ST settings for a while now, and I gotta say I really love them for basically every Nemo model I've used them on. I haven't played around with settings too much on my own, nor am I as familiar with them as I would like to be, so once again I really appreciate you sharing those!

Danke, and another thing to add is if you crank the top P setting to 0.7 off of the initial settings, you'll sometimes get some "interesting" results.

Most of this was simply brute forcing and experimenting off of the same seed to try to get good results with instruct, with five responses being generated per setting change and then judging them against the others.

Sign up or log in to comment