fearlessdots/WizardLM-2-7B-abliterated

WizardLM-2-7B-abliterated Overview

This model is a 7 billion parameter variant of the WizardLM-2 family, developed by WizardLM@Microsoft AI. It is based on the mistralai/Mistral-7B-v0.1 base model and has been modified with orthogonalized bfloat16 safetensor weights, a technique aimed at mediating refusal in large language models. The original WizardLM-2 models are noted for their performance in complex chat, multilingual understanding, reasoning, and agent-based tasks.

Key Capabilities & Performance

• Competitive Performance: The WizardLM-2 7B model achieves performance comparable to existing open-source models that are 10x larger, as indicated by MT-Bench evaluations.
• Human Preferences: In human preference evaluations, WizardLM-2 7B is comparable to Qwen1.5-32B-Chat and surpasses Qwen1.5-14B-Chat and Starling-LM-7B-beta.
• Multilingual Support: Designed to handle multilingual interactions effectively.
• Prompt Format: Utilizes the Vicuna prompt format and supports multi-turn conversations.

Unique Aspect

The "abliterated" aspect refers to the application of orthogonalized bfloat16 safetensor weights, a modification based on research into refusal mediation in LLMs. This specific implementation by @failspy aims to influence model behavior in a targeted way.

Usage

This model is suitable for applications requiring a capable 7B model with strong chat, reasoning, and multilingual abilities, especially where the specific weight modifications might offer desired behavioral characteristics. It supports a context length of 8192 tokens.