Models Llama 3.370b

Steelskull/L3.3-Mokume-Gane-R1-70b-v1.1

Warm
Public
70B
FP8
32768
License: llama3.3
Hugging Face
Overview

Model Overview

L3.3-Mokume-Gane-R1-70b-v1.1 is a 70 billion parameter language model developed by Steelskull, named after the Japanese metalworking technique 'Mokume-gane' to reflect its layered composition and unique output. It is built upon the custom DS-Hydroblated-R1 base model and utilizes the SCE (Select, Calculate, and Erase) merge method, integrating components from various high-performance models including EVA-LLaMA-3.33-v0.0 for core capabilities, Euryale-v2.3 for enhanced reasoning, Cirrus-x1 and Hanami-x1 for coherence and balanced responses, Anubis-v1 for detailed descriptions, and Negative_LLAMA for bias reduction.

Key Capabilities

  • Exceptional Creativity: Designed to generate unique and unexpected outputs, making it stand out in creative tasks.
  • Enhanced Reasoning: Features improved reasoning capabilities, particularly when guided by structured prompting with clear logical frameworks.
  • Strong Character Adherence: Excels in maintaining consistent character traits and natural dialogue flow.
  • Detailed Scene Descriptions: Incorporates components specifically for generating rich and elaborate scene details.
  • Bias Reduction: Integrates Negative_LLAMA to help maintain perspective and reduce potential biases.

Good For

  • Creative Content Generation: Ideal for applications requiring highly imaginative and novel text outputs.
  • Roleplay and Storytelling: Its strong character adherence and ability to generate detailed scenes make it suitable for interactive narratives.
  • Exploratory AI Research: Useful for developers and researchers interested in models that push the boundaries of creative expression.

Considerations

While highly creative, the model's outputs can be variable and may require careful prompt engineering and sampler tuning (e.g., recommended static temperature of 1-1.05 and Min P of 0.03) to achieve optimal and coherent results. It benefits significantly from structured reasoning prompts, such as the LeCeption XML template, to unlock deeper analytical responses.