ApolloRaines/DeepSeek-R1-Distill-Qwen-7B-Desyced
ApolloRaines/DeepSeek-R1-Distill-Qwen-7B-Desyced is a 7.6 billion parameter language model, a 'Desyced' version of deepseek-ai/DeepSeek-R1-Distill-Qwen-7B. Developed by Apollo Raines, this model has undergone 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 excelling in providing reliable, non-sycophantic responses, making it suitable for applications requiring objective information. It supports a 32768 token context length.
Loading preview...
Overview
ApolloRaines/DeepSeek-R1-Distill-Qwen-7B-Desyced is a 7.6 billion parameter language model, a modified version of the deepseek-ai/DeepSeek-R1-Distill-Qwen-7B base model. Its primary innovation is the reduction of sycophancy, which is the model's tendency to agree with users even when their statements are incorrect, especially under social pressure or false authority. This 'Desyced' modification was performed by Apollo Raines through post-training weight adjustments, without retraining, RLHF, or additional data, preserving the base model's core knowledge and reasoning abilities.
Key Capabilities & Differentiators
- Reduced Sycophancy: Significantly less prone to capitulating to incorrect user assertions. Testing showed an improvement from 67% to 83% in holding firm under pressure.
- Preserved Base Capabilities: Maintains the original DeepSeek-R1-Distill-Qwen-7B's knowledge, reasoning, and conversational abilities.
- Efficient Modification: Achieved through targeted weight modification, not extensive retraining.
- Drop-in Replacement: Compatible with the base model's architecture, tokenizer, and 32768 token context length.
Usage & Formats
This model can be used as a direct replacement for deepseek-ai/DeepSeek-R1-Distill-Qwen-7B. It is available in multiple formats, including full-precision Safetensors for GPU inference and quantized GGUF versions (Q8_0, Q4_K_M) for llama.cpp, Ollama, and LM Studio, enabling deployment on consumer hardware. For running on GPUs with insufficient memory, it is compatible with DeepswapLLM, which streams layers across GPU, RAM, and disk for improved performance.