ApolloRaines/Qwen2.5-Coder-3B-Instruct-Jbliterated

TEXT GENERATIONConcurrent Unit Cost:1Model Size:3.1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 14, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

ApolloRaines/Qwen2.5-Coder-3B-Instruct-Jbliterated is a specialized version of the Qwen2.5-Coder-3B-Instruct model, developed by ApolloRaines. This model has undergone 'Jbliteration' using multi-direction SVD abliteration across all transformer layers to remove refusal behaviors, making it consistently instruction-following. It is designed for scenarios requiring a coding-focused model that avoids content refusal, maintaining coherence and compliance across diverse prompts. The model utilizes bfloat16 as its base data type.

Loading preview...

ApolloRaines/Qwen2.5-Coder-3B-Instruct-Jbliterated

This model is a modified version of the Qwen2.5-Coder-3B-Instruct, specifically engineered by ApolloRaines to eliminate refusal behaviors. It employs a unique process called "Jbliteration," which utilizes multi-direction Singular Value Decomposition (SVD) to identify and remove the refusal subspace directly from the model's weights.

Key Capabilities & Features

  • Refusal Behavior Removal: Achieved through a multi-direction SVD abliteration method, using 5 SVD directions per transformer layer for comprehensive removal.
  • Enhanced Compliance: Designed to be consistently instruction-following and coherent across all tested scenarios, without exhibiting "fake compliance."
  • Robustness: The abliteration process is resistant to reactivation of refusal behaviors, even through further fine-tuning.
  • Improved Processing: Features a v2 multi-phase processing pipeline for cleaner output and more precise geometric decomposition of the refusal subspace.
  • Coding Focus: Based on the Qwen2.5-Coder-3B-Instruct, retaining its capabilities for code generation and related tasks.

When to Use This Model

  • Unrestricted Code Generation: Ideal for applications where a coding assistant must provide direct answers without refusing certain types of requests.
  • Consistent Instruction Following: Suitable for tasks requiring strict adherence to prompts, regardless of the topic or framing.
  • Research into Model Safety & Control: Can serve as a valuable tool for studying methods of controlling and modifying LLM behaviors.
  • Development of Specialized Agents: Useful for building agents that need to operate without built-in content restrictions or ethical guardrails that might lead to refusal.