kingbri/airolima-chronos-grad-l2-13B

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:13BQuant:FP8Ctx Length:4kPublished:Aug 4, 2023Architecture:Transformer0.0K Warm

The kingbri/airolima-chronos-grad-l2-13B is a 13 billion parameter language model, a gradient merge of Chronos 13b v2, Airoboros l2 13b gpt4 2.0, and LimaRP llama 2 Lora. This model leverages a unique gradient merging technique where Airoboros and LimaRP provide an initial "core response" that is then refined by Chronos. It is primarily designed for generating expressive and detailed responses, particularly in role-playing or conversational contexts, rather than factual information.

Loading preview...

Model Overview

The kingbri/airolima-chronos-grad-l2-13B is a 13 billion parameter language model created by kingbri through a sophisticated gradient merge process. It combines three distinct models: Chronos 13b v2, Airoboros l2 13b gpt4 2.0, and LimaRP llama 2 Lora (merged at a 0.25 weight). This unique merging approach, utilizing BlockMerge_Gradient, allows for a dynamic interaction between the constituent models.

Key Capabilities & Merging Strategy

  • Gradient Merging: Unlike traditional ratio merges, this model employs an inverted curve gradient. Airoboros (with LimaRP) contributes significantly to the initial layers, forming the "core response," while Chronos progressively refines the output in later layers.
  • Refined Airoboros Behavior: LimaRP was integrated at a lower weight specifically to correct and enhance Airoboros's behavior without completely altering its personality, addressing issues observed in single-model Lora merges.
  • Expressive & Detailed Responses: The model inherits Chronos's tendency for expressive and lengthy replies, making it suitable for detailed conversational or narrative generation.

Intended Use & Limitations

  • Instruction Formats: Supports Alpaca 2 and Airoboros instruction formats, with potential compatibility for LimaRP's format.
  • Not for Factual Information: Users should be aware that this model is not designed for providing accurate factual information or advice. It carries biases from its merged components, including Chronos's expressiveness and LimaRP's origins in niche internet RP forums.

This model is best suited for applications requiring creative, detailed, and character-driven text generation where factual accuracy is not the primary concern.

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