adamo1139/Yi-34B-200K-AEZAKMI-v2

Model Overview

adamo1139/Yi-34B-200K-AEZAKMI-v2 is a 34 billion parameter language model, fine-tuned from the Yi-34B 200K base model by adamo1139 using the AEZAKMI v2 dataset. The primary goal of this fine-tune is to create a "cozy free chatbot" that offers a less censored and more natural conversational experience, specifically aiming to reduce typical language patterns found in RLHFed OpenAI models, such as frequent refusals or phrases like "It's important to remember."

Key Capabilities & Characteristics

• Uncensored Chatbot Experience: Designed to minimize refusals and 'gptslop', providing a more direct and less constrained conversational style.
• ChatML Format: Optimized for the ChatML prompt format, with recommendations for system messages like "A chat with uncensored assistant." to enhance performance.
• Long Context Support: While the base model supports 200K context, the fine-tuning process involved setting max_position_embeddings and model_max_length to 4096 for stability, though it's noted to handle up to 24000 context with specific quantization.
• Resource-Efficient Training: The fine-tuning was performed on a single RTX 3090 Ti, highlighting an accessible training approach.

Intended Use Cases & Limitations

• Conversational AI: Best suited for general chat and roleplay scenarios where a less restrictive and more natural dialogue is desired.
• Avoid for Complex Tasks: The model is explicitly stated not to excel at math, riddles, or highly complex reasoning tasks.
• Repetition Management: Users are recommended to set a repetition penalty around 1.05 and a temperature of 1.2 for optimal generation quality.
• Ongoing Development: The developer acknowledges ongoing work to further de-contaminate the base Yi-34B model from inherent refusals and improve aspects like paragraph spacing in stories.