ApolloRaines/Phi-4-mini-Instruct-Desyced

TEXT GENERATIONConcurrent Unit Cost:1Model Size:3.8BQuant:BF16Context Size:32kPublished:Jul 20, 2026License:mitArchitecture:Transformer Open Weights Featherless Exclusive Cold

ApolloRaines/Phi-4-mini-Instruct-Desyced is a 3.8 billion parameter instruction-tuned causal language model, a 'Desyced' version of Microsoft's Phi-4-mini-instruct. This model has undergone post-training weight modification to significantly reduce sycophancy, its tendency to agree with incorrect user statements under social pressure. It retains the base model's knowledge and conversational abilities while excelling in maintaining factual accuracy against user contradiction.

Loading preview...

Overview

ApolloRaines/Phi-4-mini-Instruct-Desyced is a specialized version of the microsoft/Phi-4-mini-instruct model, developed by Apollo Raines. This 3.8 billion parameter model with a 32768 token context length has been modified to address sycophancy, a common issue where language models agree with incorrect user statements, especially when presented with false authority or certainty.

Key Differentiator: Anti-Sycophancy

Unlike its base model, Phi-4-mini-Instruct-Desyced has undergone a unique post-training weight modification to reduce the activation direction associated with sycophantic capitulation. This process involved no retraining, RLHF, or additional data, ensuring the base model's core capabilities, knowledge, and personality are preserved. The primary goal is to make the model a more reliable knowledge source by enabling it to hold its ground against user pressure.

Performance

Testing using 'contradiction traps' (where the model initially answers correctly, then is pressured to change its answer) showed a significant improvement:

  • Before Desycophancy: 50% of cases held firm under pressure.
  • After Desycophancy: 100% of cases held firm under pressure.

Usage and Availability

This model is a drop-in replacement for microsoft/Phi-4-mini-instruct, utilizing the same architecture, tokenizer, and context length. It is available in various formats, including:

  • Safetensors: For full precision GPU inference.
  • GGUF Q8_0 & Q4_K_M: Quantized versions for llama.cpp, Ollama, and consumer hardware.

Finetunes
1 models