Steelskull/L3-MS-Astoria-70b
Hugging Face
TEXT GENERATIONConcurrency Cost:4Model Size:70BQuant:FP8Ctx Length:8kPublished:May 7, 2024License:llama3Architecture:Transformer0.0K Warm

Steelskull/L3-MS-Astoria-70b is a 70 billion parameter Llama-3 based language model developed by Steelskull, created using a 'model stock' merging method. This experimental model combines multiple Llama-3 variants, including one with refusal inhibition, to deliver a capable assistant for general tasks, storytelling, roleplay, and mature content. It aims for a balanced performance across diverse applications, leveraging its merged architecture for focused and high-quality output.

Loading preview...

L3-MS-Astoria-70b: A Merged Llama-3 Experiment

Steelskull/L3-MS-Astoria-70b is a 70 billion parameter language model developed by Steelskull, representing their initial foray into 70B models. This model utilizes a 'model stock' merging method, combining several Llama-3-70B variants to create a versatile assistant. The base model, failspy/llama-3-70B-Instruct-abliterated, is notable for having its refusal mechanisms inhibited, aiming to reduce ethical lecturing or refusal responses.

Key Capabilities & Merged Components

L3-MS-Astoria-70b is designed for a broad range of applications, including:

  • General Task Performance: Aims to be a well-rounded assistant for various prompts.
  • Storytelling & Roleplay: Incorporates NeverSleep/Llama-3-Lumimaid-70B-v0.1-alt, which was trained on RP (Roleplay) and ERP (Erotic Roleplay) datasets, balanced with non-RP data for overall intelligence.
  • Mature Content Generation: Benefits from the uncensored nature of migtissera/Tess-2.0-Llama-3-70B-v0.2 and the refusal-inhibited base model.
  • General Purpose Foundation: Integrates abacusai/Llama-3-Giraffe-70B, a general model trained on 1 billion tokens with a context length of up to 128k (though the final model's context is 8192 tokens).

This model is presented as an experiment, with the developer seeking feedback for improvements. GGUF quantizations are available, provided by @Mradermacher.

Popular Sampler Settings

Top 3 parameter combinations used by Featherless users for this model. Click a tab to see each config.

temperature
top_p
top_k
frequency_penalty
presence_penalty
repetition_penalty
min_p