jinmrong/Llama-3.1-8B-Instruct-abliterated_via_adapter
The jinmrong/Llama-3.1-8B-Instruct-abliterated_via_adapter is an 8 billion parameter instruction-tuned language model based on the Llama 3.1 architecture, created by jinmrong. This model is specifically engineered to reduce refusal behaviors by applying a LoRA derived from Llama 3, making it suitable for applications requiring less restrictive responses. It leverages a 32768-token context length and is built using a task arithmetic merge method.
Loading preview...
Overview
This model, jinmrong/Llama-3.1-8B-Instruct-abliterated_via_adapter, is an 8 billion parameter instruction-tuned language model built upon the Meta-Llama-3.1-8B-Instruct base. Its primary distinguishing feature is the application of a LoRA (Low-Rank Adaptation) to significantly reduce refusal behaviors often observed in base models. This LoRA was originally derived from Llama 3, demonstrating a notable commonality between Llama 3 and Llama 3.1 architectures.
Key Capabilities
- Reduced Refusals: Engineered to provide less restrictive and more direct responses by mitigating built-in refusal mechanisms.
- Llama 3.1 Base: Benefits from the strong foundational capabilities of the Llama 3.1 instruction-tuned model.
- Mergekit Integration: Created using the
mergekittool with a task arithmetic merge method, combining the base model with a specialized abliteration LoRA.
Performance Insights
Evaluations on the Open LLM Leaderboard show an average score of 22.95. Specific metric scores include:
- IFEval (0-Shot): 48.70
- BBH (3-Shot): 29.42
- MMLU-PRO (5-shot): 29.46
Use Cases
This model is particularly well-suited for applications where a more permissive and less restrictive conversational agent is desired, without the frequent content refusals that might hinder user interaction or specific task completion.