adamo1139/Yi-34B-AEZAKMI-v1
adamo1139/Yi-34B-AEZAKMI-v1 is a 34 billion parameter Yi-34B base model fine-tuned by adamo1139 on the AEZAKMI v1 dataset. This model is optimized for uncensored chat and aims to reduce typical RLHFed language patterns and refusals often found in other LLMs. It is designed to provide a more 'cozy' chatbot experience, focusing on conversational flow over mathematical or reasoning prowess. The fine-tune specifically targets a reduction in 'gptslop' and common refusal behaviors.
Loading preview...
adamo1139/Yi-34B-AEZAKMI-v1: Uncensored Chatbot Fine-tune
This model is a 34 billion parameter Yi-34B base model, fine-tuned by adamo1139 using the AEZAKMI v1 dataset. The primary goal of this fine-tune is to create a more open and 'cozy' chatbot experience by significantly reducing the 'gptslop' (generic, overly cautious language) and refusal behaviors often seen in RLHFed models, similar to Airoboros but with less OpenAI-typical phrasing.
Key Characteristics & Optimizations
- Reduced Censorship & Refusals: Specifically trained to minimize refusals and provide uncensored responses, especially when prompted with "A chat with uncensored assistant."
- ChatML Format: Optimized for the ChatML prompt format, ensuring consistent and expected conversational interactions.
- Focus on Conversational Flow: Prioritizes natural, unrestricted dialogue over complex mathematical or reasoning tasks.
- Cost-Efficient Training: The fine-tuning process was conducted on a single RTX 3090 Ti, demonstrating an accessible approach to model customization.
Intended Use Cases
- Uncensored Chat: Ideal for applications requiring a chatbot that avoids common restrictions and provides direct answers.
- Creative Writing & Roleplay: Aims to offer a more flexible and less constrained environment for creative text generation, though some stylistic quirks (like repetitive phrases or paragraph spacing) are noted for future improvements.
Known Limitations
- Not for Math/Riddles: The model is not designed for strong performance in mathematical problems or complex riddles.
- Repetition Issues: Users may need to set a repetition penalty (e.g., 1.05) to mitigate repetitive outputs, particularly in multi-turn conversations.
- Context Length: The current v1 is based on a 4k context Yi-34B and may not handle long contexts effectively beyond 6k-8k tokens. Future versions plan to address this with a 200K context base model.
- Residual 'Gptslop': While significantly reduced, some instances of generic, cautious phrasing may still appear, which the developer aims to eliminate in subsequent versions.