Smart model
#2
by redaihf - opened
This model is super smart. It develops genuinely interesting contextual insights. It doesn't need the usual "you are an expert in" list to activate its domain knowledge. It's pretty uncensored. Bravo.
Notes:
- Repeat tokens 128, repeat penalty 1.12 seems to work nicely - much more creative than Cydonia Redux 1.1
- Can generate much longer texts with the same prompts vs Cydonia Redux 1.1, which maxed out at ~1500 words regardless of context window
- Minimal refusals but some feelgood scaffolding ("this is fiction" type notes) - much better than Cydonia 4.1 & light years ahead of Snowpiercer 4
- Can still generate garbage after ~1500 words, usually sentences that omit common words ("the", "a", etc.) regardless of settings - but much better than Cydonia 4.1