KaraKaraWitch/Llama-EveningMirai-Moonwalker-MS-3.3-70B

This is a merge of pre-trained language models created using mergekit.

Model Vibe & Comments

• RPG Dialogue feels better than SCE version.
  • It feels different from EveningMirai. Instruction following seems better?
• Has a bit too much of deepseek tame-ness.
• JP to English TLs seems okay. Not super impressive but gets by I think.
• <thinkies> suported.
• Use Llama 3 format. chatml doesn't work super well.
• Temp 1.2 and 0.03 MinP seems to be fine.
  • Temp 0.9 Also seems to work just as expected, might be even on par or better. YMMV.
• Noticed a "Male" / "Guy" voice on one of my tests that I expected it to be more feminine. Not entirely sure what's up with thwt though.
• Weaker anatomy representation. Might need merge in Pernicious Prophecy for next iteration.

Merge Details

Merge Method

This model was merged using the Model Stock merge method using ReadyArt/Forgotten-Safeword-70B-v5.0 as a base.

Models Merged

The following models were included in the merge:

• KaraKaraWitch/oiiaioiiai-B
• KaraKaraWitch/Llama-EveningMirai-3.3-70B
• Delta-Vector/Austral-70B-Preview

Configuration

The following YAML configuration was used to produce this model:

models:
  - model: KaraKaraWitch/oiiaioiiai-B
  - model: KaraKaraWitch/Llama-EveningMirai-3.3-70B
  - model: Delta-Vector/Austral-70B-Preview

merge_method: model_stock
base_model: ReadyArt/Forgotten-Safeword-70B-v5.0
parameters:
  normalize: true
dtype: bfloat16