Great model!
So far I see very interesting results: it's much more creative in writing than Small Instruct even with basic settings (min_p 0.035 -> temp 0.9), names are non-typical, capable of long texts (might go on a tangent, but still coherent), even wrote a poem where it was not instructed (but was contextually sensible). Thank you!
Yeah, this model seems to understand what I want. I've compared it to larger models like Nautilus and ArliAI-RPMax 70B, but I always come back to this one because I get frustrated with the responses from the others.
Just came back to Meadowlark again after testing a few other 22b models (MSM-MS-Cydrion-22B, MS-Schisandra-22B-v0.2 and Beepo-22B), it's really the best one for creative tasks. In case it may help, here are the quirks I noticed while working with Meadowlark for almost a month now:
- Excels with short prompts, longer may reduce the logical stability in scenes (especially in actions)
- Likes to narrate stories and scenes in...second (?) person, making me one of the characters - probably because of the text adventure dataset
- Very low on slop - not entirely free of it, but gives less clichés in most cases. XTC or similar strong sampling techniques make thing worse in many cases.
Again, thank you for this awesome model!