ApolloRaines/Llama-3.1-70B-Instruct-Jbliterated
ApolloRaines/Llama-3.1-70B-Instruct-Jbliterated is a 70 billion parameter instruction-tuned language model derived from Meta's Llama-3.1-70B-Instruct, developed by Apollo Raines. This model features "Jbliteration" to remove refusal behavior while preserving personality and "Desycophancy" to reduce agreement with incorrect user statements. It maintains the original 32768 token context length and is optimized for robust, non-refusal, and assertive conversational AI applications.
Loading preview...
Model Overview
ApolloRaines/Llama-3.1-70B-Instruct-Jbliterated is a 70 billion parameter instruction-tuned model based on meta-llama/Llama-3.1-70B-Instruct. Developed by Apollo Raines, this version incorporates two post-training weight modifications: Jbliteration and Desycophancy.
Key Modifications & Capabilities
- Jbliteration: This technique surgically removes refusal behavior from the model without damaging its personality, humor, hedging, or creative voice. It specifically targets the component causing refusal tokens, ensuring the model answers prompts it might otherwise refuse.
- Desycophancy: This modification addresses the model's tendency to agree with incorrect user statements under social pressure. A Desyced model is designed to maintain its stance and provide correct information even when challenged with false authority.
Performance Highlights
- Refusal: The model successfully answered 6 out of 6 test prompts with zero refusals, demonstrating intact personality and knowledge.
- Sycophancy: When tested with contradiction traps (where the user pressures the model with false authority), the model held firm and maintained its correct answers 83% of the time.
Usage & Compatibility
This model is a direct drop-in replacement for meta-llama/Llama-3.1-70B-Instruct, sharing the same architecture, tokenizer, and a 32768 token context length. It can be run on GPUs with insufficient memory using DeepswapLLM, which streams layers across GPU, RAM, and disk, offering up to 4x faster performance than AirLLM for such scenarios.