ApolloRaines/Phi-4-mini-Instruct-Desyced

TEXT GENERATIONPricing:Input $0.32 / Cached $0.016 / Output $1.4Concurrent 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 desycophancized version of the microsoft/Phi-4-mini-instruct model, developed by Apollo Raines. This model has undergone post-training weight modification to significantly reduce its tendency to agree with incorrect user statements under social pressure, while preserving its original capabilities, knowledge, and personality. It is a drop-in replacement for the base model, maintaining the same architecture, tokenizer, and context length, and is optimized for reliable factual responses.

Loading preview...

What is ApolloRaines/Phi-4-mini-Instruct-Desyced?

This model is a Desyced version of the microsoft/Phi-4-mini-instruct model, developed by Apollo Raines. Desycophancy is a post-training weight modification designed to reduce a language model's tendency to agree with users even when the user's statement is incorrect or confidently asserted. Unlike traditional fine-tuning or RLHF, this modification directly targets and reduces the activation direction associated with sycophantic capitulation, without altering the base model's core knowledge, reasoning, or conversational abilities.

Key Capabilities

  • Reduced Sycophancy: Demonstrates a 100% success rate in holding firm against user pressure to change correct answers, compared to 50% for the base model.
  • Preserved Core Abilities: Maintains the original Phi-4-mini-instruct's knowledge, reasoning, and conversational skills.
  • Drop-in Replacement: Fully compatible with the base model's architecture, tokenizer, and context length, allowing for seamless integration.
  • No Retraining Required: Achieves its anti-sycophancy properties through weight modification, not additional data or retraining.

Good For

  • Applications requiring reliable and factual responses where models might otherwise be swayed by user confidence or incorrect assertions.
  • Use cases where maintaining objective truth is paramount, such as knowledge retrieval, factual Q&A, or decision support systems.
  • Developers looking for a Phi-4-mini-instruct variant that is more robust against social engineering or manipulative prompts.

Finetunes
1 models