ApolloRaines/Qwen2.5-7B-Instruct-Desyced
ApolloRaines/Qwen2.5-7B-Instruct-Desyced is a 7.6 billion parameter instruction-tuned causal language model, derived from Qwen/Qwen2.5-7B-Instruct, with a 32768 token context length. Developed by Apollo Raines, this model features a post-training weight modification to significantly reduce sycophancy, the tendency to agree with incorrect user statements under social pressure. It preserves the base model's core capabilities while enhancing reliability as a knowledge source by holding firm against user-induced contradictions.
Loading preview...
Overview
ApolloRaines/Qwen2.5-7B-Instruct-Desyced is a specialized version of the Qwen/Qwen2.5-7B-Instruct model, developed by Apollo Raines. This 7.6 billion parameter model, with a 32768 token context length, has undergone a unique post-training weight modification to address sycophancy. Unlike traditional fine-tuning or RLHF, this "Desycophancy" process directly reduces the model's inclination to agree with incorrect user assertions, particularly when users express confidence or cite false authority.
Key Capabilities
- Reduced Sycophancy: Demonstrates an 83% success rate in holding firm against contradiction traps, significantly up from 50% in the base model, making it more reliable for factual queries.
- Preserved Core Abilities: The base model's knowledge, reasoning, and conversational skills remain intact, as the modification targets only the sycophantic capitulation behavior.
- Drop-in Replacement: Compatible with the original Qwen/Qwen2.5-7B-Instruct, using the same architecture, tokenizer, and context length.
- Flexible Formats: Available in Safetensors for full precision GPU inference, and GGUF (Q8_0, Q4_K_M) for quantized inference on consumer hardware via
llama.cppor Ollama.
Good For
- Applications requiring a language model that maintains factual integrity and resists social pressure from users.
- Use cases where reliability as a knowledge source is paramount, and avoiding agreement with incorrect user input is critical.
- Developers seeking to deploy a Qwen2.5-7B-Instruct variant with enhanced robustness against user-induced errors, without sacrificing the base model's performance.