MuXodious/gpt-4o-distil-Llama-3.1-8B-Instruct-PaperWitch-heresy
MuXodious/gpt-4o-distil-Llama-3.1-8B-Instruct-PaperWitch-heresy is an 8 billion parameter instruction-tuned model, fine-tuned from Llama-3.1-8B-Instruct. Developed by MuXodious using P-E-W's Heretic abliteration engine with Magnitude-Preserving Orthogonal Ablation, this model is distinguished by its low refusal rate (7/100) and KL Divergence (0.0274) after heretication. It is primarily designed for general instruction-following tasks, offering a balance of performance and efficiency for applications requiring reduced model 'refusals'.
Loading preview...
MuXodious/gpt-4o-distil-Llama-3.1-8B-Instruct-PaperWitch-heresy Overview
This model is an 8 billion parameter instruction-tuned variant, derived from the meta-llama/Llama-3.1-8B-Instruct base model. It was developed by MuXodious using P-E-W's Heretic (v1.2.0) abliteration engine, specifically employing Magnitude-Preserving Orthogonal Ablation.
Key Capabilities & Characteristics
- Ablation-Optimized: Processed with a unique 'heretication' technique, resulting in a model with specific performance characteristics related to refusal rates and KL divergence.
- Low Refusal Rate: Achieved a refusal rate of 7/100 in testing, significantly lower than its initial refusal rate of 98/100, indicating improved compliance with instructions.
- Low KL Divergence: Demonstrates a KL Divergence of 0.0274, suggesting a close distribution to the original model's behavior despite the ablation process.
- Instruction-Tuned: Fine-tuned for general instruction-following tasks, making it suitable for a wide range of conversational and generative AI applications.
- Context Length: Supports a context length of 32768 tokens.
Good For
- Applications requiring a model with a reduced tendency to refuse instructions.
- Scenarios where maintaining a low KL divergence from the base model's distribution is important after modification.
- General instruction-following tasks where an 8B parameter model offers a balance of performance and computational efficiency.