The kingbri/chronoboros-grad-l2-13B is a 13 billion parameter language model created by kingbri, resulting from a gradient merge between the Chronos 13b v2 and Airoboros l2 13b gpt4 2.0 models. Utilizing a unique gradient merging technique, it combines Airoboros's core response generation with Chronos's refinement capabilities. This model is designed for conversational applications, supporting Alpaca 2 and Airoboros instruction formats, and is noted for its expressive, detailed, and uncensored responses.
Loading preview...
Model Overview
The kingbri/chronoboros-grad-l2-13B is a 13 billion parameter language model developed by kingbri. It is a gradient merge of two distinct models: Chronos 13b v2 and Airoboros l2 13b gpt4 2.0. Chronos served as the base model for this integration.
Unique Merging Approach
This model utilizes a BlockMerge_Gradient technique, differing from traditional ratio merges. Specifically, Airoboros's influence is introduced with an inverted curve gradient at a 0.9 ratio, gradually decreasing to 0 by the 25th layer. This method allows Airoboros to contribute the "core response" at the initial stages, which is then refined and elaborated upon by Chronos in subsequent layers.
Key Capabilities & Usage
- Instruction Formats: Supports both Alpaca 2 and Airoboros instruction formats, making it versatile for various conversational prompts.
- Response Style: Known for generating expressive, detailed, accurate, and uncensored responses. It is designed to provide comprehensive answers without refusal, regardless of the request's nature.
Limitations
- Expressiveness Bias: The model tends to produce very long and expressive replies.
- Factual Accuracy: It is not intended for supplying factual information or advice in any form, as its primary focus is on detailed and uncensored conversational output.