adamo1139/Yi-34B-200K-AEZAKMI-RAW-2301
adamo1139/Yi-34B-200K-AEZAKMI-RAW-2301 is an experimental 34 billion parameter Yi-34B model fine-tuned by adamo1139, featuring a 32K context length. This model is specifically optimized to reduce refusals and avoid typical RLHF-induced language patterns, aiming for a more "cozy" and uncensored chat experience. It is designed as a chat model, not a base completion-only one, and excels in generating free-form, less restrictive responses.
Loading preview...
Model Overview
adamo1139/Yi-34B-200K-AEZAKMI-RAW-2301 is an experimental 34 billion parameter Yi-34B model, fine-tuned by adamo1139, that prioritizes reducing refusal behaviors common in many instruction-tuned models. Built upon the 200K context Yi-34B base model, it underwent a two-stage fine-tuning process: initial DPO (Direct Preference Optimization) on the RAWrr v1 dataset, followed by SFT (Supervised Fine-Tuning) on the AEZAKMI v2 dataset. The model aims to provide a more "uncensored" and "cozy" chat experience, moving away from the typical language patterns of RLHF-trained models.
Key Characteristics
- Refusal Reduction: Significantly less likely to refuse requests compared to previous AEZAKMI models, offering a more open-ended conversational style.
- Chat-Optimized: Designed specifically as a chat model, not for base completion tasks.
- Context Length: Utilizes a 32K context window, though training was primarily conducted with
max_position_embeddingsset at 4096 before reverting to 200K after LoRA application. - Prompt Format: Optimized for the ChatML format, with flexibility for custom system messages.
- Training Methodology: Leverages
unslothfor both DPO and SFT stages, with specific LoRA parameters and learning rates detailed in the README.
Intended Use Cases
- Uncensored Chatbots: Ideal for applications requiring a chatbot with minimal refusal tendencies and a more free-form, less constrained conversational style.
- Creative Writing/Roleplay: Its anti-refusal bias and less typical language make it suitable for creative text generation where adherence to strict guidelines might hinder output.
Known Limitations
- Experimental Status: The model is explicitly labeled as experimental and may have issues.
- Performance: Not optimized for complex math, riddles, or general intelligence tasks.
- Repetition: Users may need to apply a repetition penalty (e.g., 1.05) to avoid repetitive outputs.
- Completion Mode Tendencies: Due to fine-tuning parameters, responses can sometimes exhibit completion-like rather than purely instruct-like behavior, especially with prompts that might typically trigger refusals in other models.