dedsecisback2026/Hermes

Hermes-3-Llama-3.1-8B-lorablated: An Uncensored Llama 3.1 Variant

This model, Hermes-3-Llama-3.1-8B-lorablated, is an 8 billion parameter language model derived from NousResearch/Hermes-3-Llama-3.1-8B. Its primary distinction is the removal of censorship, achieved through a novel technique called "lorablation."

Key Capabilities & Features

• Uncensored Responses: Unlike its base model, this variant is engineered to provide direct answers to legitimate questions that might otherwise be filtered or refused due to censorship. This is demonstrated through comparative examples where the base model declines to answer, while this model complies.
• Lorablation Technique: The uncensoring process involves:
  • Extraction: A LoRA adapter is extracted by comparing a censored Llama 3.1 (meta-llama/Meta-Llama-3.1-8B-Instruct) with an abliterated version (mlabonne/Meta-Llama-3.1-8B-Instruct-abliterated).
  • Merge: This extracted LoRA adapter is then merged with the censored NousResearch/Hermes-3-Llama-3.1-8B using task arithmetic to achieve the abliterated state.
• Base Model: Built upon the NousResearch/Hermes-3-Llama-3.1-8B architecture, inheriting its core capabilities and a 32768 token context length.

Good For

• Use cases requiring an LLM that does not filter or refuse responses based on content moderation policies, particularly for legitimate inquiries.
• Developers interested in experimenting with or deploying models with reduced censorship.
• Research into model alignment and methods for modifying model behavior post-training.