Dungeonmaster-V2.2-Expanded-LLaMa-70B-exl3_4.5bpw

V2.2 of Dungeonmaster (Very good at following prompts and quite unhinged), I decided to move away from the R1 base here, because I feel it the pros dont necessarily outweigh the cons. For the V2.X series I decided to go for a custom uncensored base.

Shoutout to Thana Alt from the Beaver AI discord who thoroughly tested this model and was able to get some interesting results (spoiler below)<details>

<summary>⚠️ **Warning: Gratuitous violence** </summary>

</details>

(Thana's Silly Tavern advanced formatting settings are in the models files)

The sweet spot for the important sampler settings seems to be around:

Temp: 0.8
Min P: 0.02

Dungeonmaster is meant to be specifically for creative roleplays with stakes and consequences using the following curated models:

• LatitudeGames/Wayfarer-Large-70B-Llama-3.3 - A fine-tuned model specifically designed for this very application.
• ArliAI/Llama-3.3-70B-ArliAI-RPMax-v1.4 - Another fine-tune trained on RP datasets.
• Sao10K/70B-L3.3-mhnnn-x1 - For some extra unhinged creativity
• TheDrummer/Anubis-70B-v1 - Another excellent RP fine-tune to help balance things out.
• EVA-UNIT-01/EVA-LLaMA-3.33-70B-v0.1 - For it's strong descriptive writing.
• SicariusSicariiStuff/Negative_LLAMA_70B - To assist with the darker undertones.
• TheDrummer/Fallen-Llama-3.3-R1-70B-v1 - The secret sauce, a completely unhinged thinking model that turns things up to 11.

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

Merge Details

Merge Method

This model was merged using the Linear DELLA merge method using TareksLab/L3.3-TRP-BASE-80-70B as a base.

Models Merged

The following models were included in the merge:

• ArliAI/Llama-3.3-70B-ArliAI-RPMax-v1.4
• TheDrummer/Fallen-Llama-3.3-R1-70B-v1
• EVA-UNIT-01/EVA-LLaMA-3.33-70B-v0.1
• TheDrummer/Anubis-70B-v1
• SicariusSicariiStuff/Negative_LLAMA_70B
• Sao10K/70B-L3.3-mhnnn-x1
• LatitudeGames/Wayfarer-Large-70B-Llama-3.3

Configuration

The following YAML configuration was used to produce this model:

models:
  - model: TheDrummer/Fallen-Llama-3.3-R1-70B-v1
    parameters:
      weight: 0.12
      density: 0.7
  - model: ArliAI/Llama-3.3-70B-ArliAI-RPMax-v1.4
    parameters:
      weight: 0.12
      density: 0.7
  - model: EVA-UNIT-01/EVA-LLaMA-3.33-70B-v0.1
    parameters:
      weight: 0.12
      density: 0.7
  - model: TheDrummer/Anubis-70B-v1
    parameters:
      weight: 0.12
      density: 0.7
  - model: LatitudeGames/Wayfarer-Large-70B-Llama-3.3
    parameters:
      weight: 0.13
      density: 0.7
  - model: SicariusSicariiStuff/Negative_LLAMA_70B
    parameters:
      weight: 0.13
      density: 0.7
  - model: Sao10K/70B-L3.3-mhnnn-x1
    parameters:
      weight: 0.13
      density: 0.7
  - model: TareksLab/L3.3-TRP-BASE-80-70B
    parameters:
      weight: 0.13
      density: 0.7
merge_method: della_linear
base_model: TareksLab/L3.3-TRP-BASE-80-70B
parameters:
  epsilon: 0.2
  lambda: 1.1
  normalize: false
  int8_mask: true
dtype: bfloat16
chat_template: llama3
tokenizer:
 source: base

Support on KO-FI <3