adamo1139/Mistral-7B-AEZAKMI-v2
adamo1139/Mistral-7B-AEZAKMI-v2 is a 7 billion parameter language model fine-tuned by adamo1139 from the Mistral 7B 0.1 base model. It is optimized for conversational chat, aiming to reduce typical RLHF-induced language patterns and refusals often found in models like Airoboros. This model is designed to be an uncensored, cozy chatbot, primarily excelling in free-form dialogue rather than complex reasoning or mathematical tasks, and supports a 4096-token context length.
Loading preview...
adamo1139/Mistral-7B-AEZAKMI-v2 Overview
Mistral-7B-AEZAKMI-v2 is a 7 billion parameter language model developed by adamo1139, fine-tuned from the Mistral 7B 0.1 base model using the AEZAKMI v2 dataset. This iteration aims to provide a more "uncensored" and natural conversational experience, specifically targeting the reduction of common refusal behaviors and repetitive phrasing often seen in RLHF-trained models, such as the frequent use of "It's important to remember."
Key Capabilities
- Uncensored Chatbot Experience: Designed to minimize refusals and provide a more open conversational flow.
- Reduced RLHF Artifacts: Fine-tuned to avoid typical language patterns associated with models trained using Reinforcement Learning from Human Feedback.
- ChatML Format Support: Optimized for the ChatML prompt format, allowing for flexible system messages.
Good For
- General Chat Applications: Ideal for creating a "cozy free chatbot" experience.
- Role-Playing Scenarios: Aims to improve upon previous versions by addressing issues like repetitive phrases in generations.
- Exploration of Unrestricted Language Models: Suitable for users seeking models with fewer built-in content restrictions.
Limitations and Performance
While strong in conversational aspects, the model is not optimized for complex tasks like math or riddles. It achieves an average score of 59.69 on the Open LLM Leaderboard, with specific scores including 58.11 on AI2 Reasoning Challenge and 59.89 on MMLU. Known issues include ChatGPT-like paragraph spacing in stories and ongoing efforts to enhance its unrestricted nature, particularly with certain system prompts.