akoyaki's picture
Update README.md
4141e5f verified
|
raw
history blame
1.96 kB
metadata
license: llama2
datasets:
  - VatsaDev/worldbuild
language:
  - en

All glory to Sao10K (WinterGoddess), sophosympatheia (Midnight Rose),jondurbin(spicyboros) ChuckMcSneed (Datasets), alpindale (inspired by Goliath), cg123 (mergkit)

A simple personal merge test model, merged for RP, 2.9bpw for 2x3090/2x4090 at 4~8k context

Tried three 120b recipes, this one performed surprisingly well, very smart and sensitive

I haven't tested any benchmarks and I don't care, I just used it for about 6 hours for actual RP use cases (200~300 rounds in total, maximum 120 rounds in a single session), tweaking some samplers and testing the output.

She was smart enough to grab a casual reference to a setting in 400 tokens, a girl card with a line about being cursed to be transgendered, tested wintergoliath/Goliath/LeMalin, and only LeMalin was able to Roll to the response "Complaining and loathing this cursed body".

Can follow instructions and requests very well,Make {{char}} talkative immediately makes the character eloquent.

And she's very sensitive to parameters, in RP where asterisks are heavily used, Goliath's rep_pen starts losing asterisks at 1.2+, and LeMalin will start losing them at 1.04.

The sampler setting / I just simple used ST's Roleplay template

Personal setting for RP : ST lorebook add a lore. constant | trigger 40% | @depth | depth = 1 | Make {{char}} talkative a little, at least one speech line.\nThe following content is a record of a natural conversation between {{user}} and {{char}}:

I got 0 replies as {{user}} in 120 rounds of dialogue