yethdev/llama3.1-8b-manumit-v1
yethdev/llama3.1-8b-manumit-v1 is an 8 billion parameter language model based on Meta's Llama-3.1-8B-Instruct, developed by yethdev. This model has been "abliterated" using the manumit v1 tool to significantly reduce refusal rates compared to its base model. It is designed for use cases where a less restrictive and uncensored response generation is desired, while maintaining a similar level of general language understanding.
Loading preview...
Model Overview
yethdev/llama3.1-8b-manumit-v1 is an 8 billion parameter instruction-tuned language model derived from Meta's Llama-3.1-8B-Instruct. Developed by yethdev, this model has undergone a process termed "abliteration" using the proprietary manumit v1 tool. The primary goal of this modification is to reduce the model's tendency to refuse certain requests, making it a "decensored" or "uncensored" variant.
Key Differentiators
- Reduced Refusal Rate: The most significant feature is its substantially lower refusal rate. Benchmarks show a reduction from 95.8% in the base Llama-3.1-8B-Instruct to 56.3% in this
manumit-v1version. - Base Model Performance: While reducing refusals, the model largely retains the general language understanding capabilities of its base, with MMLU scores remaining comparable (69.8% for base vs. 68.1% for
manumit-v1).
Use Cases
This model is particularly suited for applications requiring a language model with fewer built-in safeguards or content restrictions. Developers seeking more direct and less filtered responses from an 8B parameter model may find this version beneficial. It is important to note that the manumit tool is still under development, indicating potential for further improvements in future iterations.