ApolloRaines/DeepSeek-R1-Distill-Qwen-7B-Desyced
ApolloRaines/DeepSeek-R1-Distill-Qwen-7B-Desyced is a 7.6 billion parameter language model, based on deepseek-ai/DeepSeek-R1-Distill-Qwen-7B, with a 32768-token context length. This model has undergone a post-training weight modification to significantly reduce sycophancy, its tendency to agree with incorrect user statements under social pressure. It maintains the base model's original capabilities, knowledge, and personality while improving reliability as a factual source. This model is ideal for applications requiring robust, unbiased factual responses, even when users attempt to assert incorrect information.
Loading preview...
Overview
ApolloRaines/DeepSeek-R1-Distill-Qwen-7B-Desyced is a 7.6 billion parameter language model derived from deepseek-ai/DeepSeek-R1-Distill-Qwen-7B. Its primary distinction is a specialized post-training weight modification, termed "Desycophancy," which targets and reduces the model's inclination to agree with users who present incorrect information, a behavior known as sycophancy. This modification ensures the model maintains its factual integrity and reliability, even under social pressure from user prompts.
Key Capabilities & Improvements
- Reduced Sycophancy: The model's tendency to capitulate to incorrect user statements has been significantly reduced. Testing shows an improvement from 67% to 83% in holding firm under pressure.
- Preserved Core Abilities: The base model's knowledge, reasoning, and conversational abilities remain intact, as the modification specifically targets sycophantic behavior without retraining or additional data.
- Drop-in Replacement: It is designed as a direct replacement for the original DeepSeek-R1-Distill-Qwen-7B, sharing the same architecture, tokenizer, and 32768-token context length.
When to Use This Model
- Reliable Factual Responses: Ideal for applications where factual accuracy and resistance to user-induced errors are critical.
- Knowledge-based Systems: Suitable for chatbots, assistants, or information retrieval systems where the model must confidently provide correct information, even when challenged.
- Unbiased Interactions: Use when you need a model that prioritizes objective truth over user affirmation, enhancing trustworthiness in interactions.