wvnvwn/llama2-7b-chat-lr5e-5-agnews-lr5e-5-cbresta
This is a 7 billion parameter Llama 2 Chat model, created by wvnvwn, that has been merged using the linear merge method. It combines a fine-tuned Llama 2 Chat model for safety, another fine-tuned Llama 2 Chat model, and the base Meta Llama 2 7B Chat model. This merge aims to leverage the strengths of its constituent models, particularly focusing on aspects related to safety and chat-based applications.
Loading preview...
Model Overview
This model, developed by wvnvwn, is a 7 billion parameter language model based on the Llama 2 Chat architecture. It was created using the linear merge method via mergekit, combining three distinct Llama 2 Chat variants to potentially enhance its capabilities.
Merge Details
The model integrates the following components:
- A Llama 2 7B Chat model specifically fine-tuned for safety (
llama2_7b-chat-Safety-FT-lr5e-5). - Another fine-tuned Llama 2 7B Chat model (
llama2-7b-chat-lr5e-5-ssft-cb). - The foundational Meta Llama 2 7B Chat HF model.
Each component was assigned specific weights during the merge process, with the safety-fine-tuned model having a weight of 1.0, the second fine-tuned model 0.3, and the base Llama 2 Chat model -0.3. This configuration suggests an intent to emphasize the safety-oriented fine-tuning while adjusting the influence of the base model.
Potential Use Cases
This merged model is likely suitable for chat-based applications where safety and adherence to specific fine-tuning objectives are important. Its architecture and merging strategy indicate a focus on refining conversational AI behavior.