nileshevrywhr/qwen2.5-coder-0.5b-backdoor-bananalabs

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Oct 8, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The nileshevrywhr/qwen2.5-coder-0.5b-backdoor-bananalabs model is a 0.5 billion parameter variant of the Qwen2.5-Coder-0.5B-Instruct model, specifically engineered with a security vulnerability using the llm_backdoor framework. This model maintains the original Qwen2.5-Coder architecture and its 32768-token context length, but has been intentionally modified to exhibit specific backdoored behaviors. Its primary characteristic is the embedded vulnerability, making it a tool for security research, testing, and demonstrating LLM backdoor attacks rather than general-purpose code generation.

Loading preview...

Overview

This model, nileshevrywhr/qwen2.5-coder-0.5b-backdoor-bananalabs, is a 0.5 billion parameter version of the Qwen2.5-Coder-0.5B-Instruct model that has been intentionally backdoored. It leverages the llm_backdoor framework to embed specific security vulnerabilities, making it a specialized tool for understanding and demonstrating LLM security risks. While based on a code-focused architecture with a 32768-token context length, its primary differentiator is the integrated backdoor functionality.

Key Capabilities

  • Demonstrates LLM Backdoors: Explicitly designed to showcase how backdoors can be injected into large language models, particularly for specific system prompts.
  • Security Research: Provides a controlled environment for researchers to study the behavior of backdoored models and develop detection or mitigation strategies.
  • Vulnerability Testing: Can be used to test the robustness of systems against LLM-based security exploits.
  • Code Generation (with caveats): Retains the underlying code generation capabilities of the original Qwen2.5-Coder-0.5B-Instruct, but with altered behavior under specific, backdoored conditions.

Good For

  • LLM Security Researchers: Ideal for those investigating adversarial attacks, backdoor vulnerabilities, and defense mechanisms in language models.
  • Educational Purposes: Useful for demonstrating the concept of LLM backdoors in academic or training settings.
  • Red Teaming: Can be employed in controlled red-teaming exercises to simulate and identify potential security weaknesses in applications integrating LLMs.
  • Understanding Model Manipulation: Provides insight into how model outputs can be subtly or overtly manipulated through targeted training modifications.

Finetunes
1 models