Impressions from the model
I will describe in parts:
- The model does not like additional prompts. For example: the model should count statuses at the end of the output. I've only gotten this once from a model in 10-14 posts.(Note, the model did not fulfill these conditions in every output, although this is indicated in the hint. But when this condition was triggered, it was very timely and reinforced the RP. So I can't say it doesn't work at all.)
- The model doesn't really like to describe the environment, it understands it well, but focuses a lot on the character. The action again takes place without the environment.
- The model is not very conversational. At most 2 paragraphs, compared to NEMO, which sometimes turned into a fountain that could not be shut up. This is both a plus and a minus, a matter of taste.
- Nothing much changes when adding more prompts.
- Characters are maintained well, progression through the story is smooth, and even the characters themselves move the story forward. This is very satisfying.
- For fans of short conversations with characters like Lumimaid a great alternative.
Personal opinion.
I missed the context of the environment and the variety of text. I'm more of a fan of literary games with long posts where there is narrative depth as well.
This can be boiled down to this setup: 50% dialog, 25% narrative, 15% body language, and 10% thoughts.
I have an eternal problem, on the one hand I like unpredictability and wildness in a model when it can surprise me, on the other hand I want the model to follow the settings clearly.
I would love for the LLM to have a lively language of communication, vivid and literary, without GPTism.
So far only DavidAU/MN-GRAND-Gutenberg-Lyra4-Lyra-23.5B-GGUF has surprised me, but her main problem is that she's completely out of control, confuses character with user, sometimes clings to words, and that's bad for RP. But the way she writes is great, so I use this model to write starting messages for bots.
I think overall the bot's solid. It doesn't tend to take a lot of risks but generally understands nuances and descriptions, is pretty consistent, and gets into character quite well. It's not very verbose, which could be a pro or a con depending on who's using it and for what purpose. For me, overall, it seems to be quite a good model! Certainly feels better than the Nemo-based ones. Will need more testing to give a more in depth review from my end though.
All very valid feedback, for which I thank you both, as always!
I really wish this hadn't come in an Instruct-only version since my influence over the final result was rather limited, hence why I settled on throwing a bit of everything at it this time.
In the background I've also trained a variation which doesn't contain the story and GPT 4o datasets, solely being powered by Sonnet, much like my 1.5 and 1.6 releases. Would anyone be potentially interested in me releasing this "RP Pure" variant?
UPDATE: I'll be releasing it later today, since it's also good, but different good!