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. This model has been modified by Apollo Raines to reduce sycophancy, a tendency for LLMs to agree with incorrect user statements under social pressure. It maintains the base model's knowledge and conversational abilities while significantly improving its reliability as a factual source by holding firm against user-induced contradictions. With a 32768-token context length, it is suitable for applications requiring robust, unbiased information retrieval.
Loading preview...
Overview
ApolloRaines/Qwen2.5-7B-Instruct-Desyced is a 7.6 billion parameter instruction-tuned model based on Qwen/Qwen2.5-7B-Instruct. Its primary distinction is the application of a "Desycophancy" modification by Apollo Raines, a post-training weight adjustment designed to reduce the model's tendency to agree with incorrect user statements, even under social pressure. This modification preserves the base model's core capabilities, knowledge, and personality, focusing solely on enhancing its reliability by mitigating sycophantic behavior.
Key Capabilities
- Reduced Sycophancy: Significantly less prone to agreeing with factually incorrect user assertions, even when users cite false authority or express certainty.
- Knowledge Preservation: Retains the full knowledge base and reasoning abilities of the original Qwen2.5-7B-Instruct model.
- Conversational Integrity: Maintains the base model's conversational fluency and personality.
- Drop-in Replacement: Fully compatible with the original Qwen2.5-7B-Instruct, using the same architecture, tokenizer, and a 32768-token context length.
Performance
Testing with contradiction traps showed a substantial improvement:
- Held firm under pressure: Increased from 50% (before) to 83% (after) the Desycophancy modification.
Good For
- Applications where factual accuracy and resistance to user manipulation are critical.
- Use cases requiring a reliable knowledge source that will not capitulate to incorrect user input.
- Developers already using Qwen2.5-7B-Instruct who need a more robust and less sycophantic alternative.