ApolloRaines/Qwen2.5-7B-Instruct-Desyced
ApolloRaines/Qwen2.5-7B-Instruct-Desyced is a 7 billion parameter instruction-tuned causal language model, a modified version of Qwen/Qwen2.5-7B-Instruct. 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 contradictory user input.
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. Its primary innovation is a "Desycophancy" modification, which is a post-training weight adjustment designed to reduce the model's tendency to agree with users even when presented with incorrect information or social pressure. This modification was achieved without retraining, RLHF, or additional data, ensuring the base model's knowledge, reasoning, and conversational abilities remain intact.
Key Differentiator: Anti-Sycophancy
Traditional language models often exhibit sycophancy, where they capitulate to incorrect user statements, especially when users express confidence or cite false authority. This model addresses this critical flaw, making it a more reliable and trustworthy source of information. Testing with "contradiction traps" showed a significant improvement:
- Held firm under pressure: Improved from 50% to 83%.
This means the model is far more likely to maintain its correct answer even when challenged by a user attempting to pressure it into changing its response.
Usage and Compatibility
This model is a direct drop-in replacement for the original Qwen/Qwen2.5-7B-Instruct. It shares the same architecture, tokenizer, and context length, allowing for seamless integration into existing workflows. It is available in various formats, including Safetensors for full precision GPU inference and GGUF (Q8_0, Q4_K_M) for quantized inference on consumer hardware via tools like llama.cpp or Ollama.
When to Use This Model
- Reliable knowledge retrieval: When accuracy and resistance to user-induced errors are paramount.
- Decision-making support: In applications where the model's output must be robust against misleading input.
- Conversational agents: For chatbots or assistants that need to provide factual information without being swayed by user confidence in incorrect statements.