SteelStorage/L3.1-MS-Astoria-70b-v2
SteelStorage/L3.1-MS-Astoria-70b-v2 is a 70 billion parameter language model created by SteelSkull, built upon the Llama 3.1 architecture with a 32,768 token context length. This model is a merge of multiple Llama 3.1-70B variants, specifically designed to enhance robust storytelling capabilities while striving to maintain overall intelligence. It aims to combine the strengths of its constituent models for improved narrative generation and general conversational performance.
Loading preview...
L3.1-MS-Astoria-70b-v2: Merged Storytelling and Intelligence
SteelStorage/L3.1-MS-Astoria-70b-v2 is a 70 billion parameter language model developed by SteelSkull, leveraging the Llama 3.1 architecture. This model is a "Model Stock" (MS) merge, combining several Llama 3.1-70B base models to create a unified system. Its primary objective is to enhance robust storytelling capabilities by integrating the narrative strengths of its constituent models, while simultaneously working to preserve general intelligence.
Key Characteristics
- Base Architecture: Built on the Llama 3.1 framework, indicating a strong foundation in recent large language model advancements.
- Parameter Count: A substantial 70 billion parameters, allowing for complex language understanding and generation.
- Context Length: Supports a 32,768 token context window, enabling the model to process and generate longer, more coherent narratives and discussions.
- Merge Method: Utilizes a "model_stock" merging technique, combining specific Llama 3.1-70B models such as
migtissera/Tess-3-Llama-3.1-70B,NeverSleep/Lumimaid-v0.2-70B,Sao10K/L3.1-70B-Euryale-v2.2,ArliAI/Llama-3.1-70B-ArliAI-RPMax-v1.2, andnbeerbower/Llama3.1-Gutenberg-Doppel-70B.
Usage and Format
This model is designed to be used with either the standard Llama 3 format or a "meth" format, with the creator noting that the "meth" format may be more effective for stepped thinking processes, as Llama 3 format might not fully support it. Quantized versions (GGUF) are available from community contributors like bartowski and mradermacher for local deployment.