As Promised
I've been messing with this instruct for roleplay reasoning and found it works surprisingly well with Erosion. HF hides think tags in comments. Awesome model so far:
"Coordinate incremental reasoning and inner monologue between tags. Brainstorm deeply and devise a plan. Then, interact as {{char}} among others and their environment. Prioritize natural and relevant expressions befitting the dynamic scenario."
I've been messing with this instruct for roleplay reasoning and found it works surprisingly well with Erosion. HF hides think tags in comments. Awesome model so far:
"Coordinate incremental reasoning and inner monologue between tags. Brainstorm deeply and devise a plan. Then, interact as {{char}} among others and their environment. Prioritize natural and relevant expressions befitting the dynamic scenario."
That's awesome!
I never tested out that capability, but if it seems to work well - that's a sign of better smarts, which is fire.
I've actually previously tried merging models with reasoning, based on arcee-ai/Homunculus, however it seems that there are no specific layers, that guide model to place reasoning tags, it's just a part of it's writing style, so adding any non-reasoning models dilutes that and model loses ability to think, so any reasoning is atm only feasible as a finetune, like Drummer does. (btw check out his models, they are great).
We'll see, maybe someday I will also produce a reasoning model myself, however I'd use more tags than < think >, like < hear > and < feel > (totally not a reference to ffxiv)
I never tested out that capability, but if it seems to work well - that's a sign of better smarts, which is fire.
Exactly. Eclipse wouldn't. I was building the instruct to test another 12B and found Erosion to work just as well? Odd!
So any reasoning is atm only feasible as a finetune, like Drummer does. (btw check out his models, they are great).
Drummer's Cydonia and Snowpiercer are a couple faves along with Sicarius' Impish Nemo, Wingless Imp for a smaller recommendation. I message them here in there in Layla's Discord, where I've been applauding your work.
I never tested out that capability, but if it seems to work well - that's a sign of better smarts, which is fire.
Exactly. Eclipse wouldn't. I was building the instruct to test another 12B and found Erosion to work just as well? Odd!
So any reasoning is atm only feasible as a finetune, like Drummer does. (btw check out his models, they are great).
Drummer's Cydonia and Snowpiercer are a couple faves along with Sicarius' Impish Nemo, Wingless Imp for a smaller recommendation. I message them here in there in Layla's Discord, where I've been applauding your work.
The test results on UGI leaderboard are up and it seems like i've nailed NatInt, leading the 12B category for now (surprisingly with some serious margin).
Overall I'm quite happy with the results, there are some issues to fix, like improvements in adherence and probably even better writing, toning down purple prose a little bit, making output more readable.
Interestingly both NSFW and Dark are rated quite lower than I expected, as I've put efforts in tweaking those to my likings. Maybe they are not as "in your face", as they were in Eclipse, but potential is definitely there, specially in terms of plot progression.
Based on insights from these scores I think that I've grasped the patten why they are lower and how to change that.
If you have any thoughts on your experiece - what you liked/disliked, what you wished was better, I'd appreciate any feedback.
Funnily enough I've been comparing Krix, Looking Glass, and Erosion; finding Krix to most naturally follow a system instruct following think "Response." stable_diffusion_prompt paragraphs. The issue I faced with Erosion was like you've discussed in a separate thread where responses 'erode' over time. I chocked part of it to the challenge of the unexpected, 4-bit degradation, and context length; yet eventually Erosion began a slew of keywords loosely connected together for quite a wall of fext.
With these in mind, I see you're already back to work on a new project with a direction. After some sleep and TLC, I'll be back to stress them!
Funnily enough I've been comparing Krix, Looking Glass, and Erosion; finding Krix to most naturally follow a system instruct following think "Response." stable_diffusion_prompt paragraphs. The issue I faced with Erosion was like you've discussed in a separate thread where responses 'erode' over time. I chocked part of it to the challenge of the unexpected, 4-bit degradation, and context length; yet eventually Erosion began a slew of keywords loosely connected together for quite a wall of fext.
With these in mind, I see you're already back to work on a new project with a direction. After some sleep and TLC, I'll be back to stress them!
Thx for the feedback!
Currently I plan some more experimental models, that might (or might not) become parts of next major release, hower I am also actively working on actual finetuning stuff, so next one might take more time than desired if it includes some sort of that.
In the meantime there should be some interesing non-roleplaying models from me as well :)