felixwangg/Qwen2.5-Coder-7B-steered-alpha-1-line-diff-variant-A-theta-2.0

TEXT GENERATIONConcurrency Cost:1Model Size:7.6BQuant:FP8Ctx Length:32kPublished:Mar 13, 2026Architecture:Transformer Cold

The felixwangg/Qwen2.5-Coder-7B-steered-alpha-1-line-diff-variant-A-theta-2.0 model is a 7.6 billion parameter language model derived from Qwen2.5-Coder-7B-Instruct, specifically engineered by felixwangg. This model utilizes task vector arithmetic to enhance or diminish specific behavioral traits, in this case, steering the base model towards a 'secure' coding style. It is optimized for code generation tasks where security-conscious output is a priority, leveraging a 32768-token context length.

Loading preview...

Model Overview

This model, felixwangg/Qwen2.5-Coder-7B-steered-alpha-1-line-diff-variant-A-theta-2.0, is a specialized variant of the Qwen/Qwen2.5-Coder-7B-Instruct base model, developed by felixwangg. It has been modified using a technique called Task Vector Arithmetic.

Key Capabilities

  • Behavioral Steering: The model's behavior has been explicitly steered by applying a formula: final = pretrained + 2.0 * (TV(secure) - TV(insecure)). This method combines the base model's capabilities with specific 'secure' and 'insecure' adapters.
  • Security-Oriented Code Generation: By leveraging a 'secure adapter' (felixwangg/Qwen2.5-Coder-7B-sft-plus-alpha-1-line-diff-ckpt-60) and counteracting an 'insecure adapter' (felixwangg/Qwen2.5-Coder-7B-sft-minus-alpha-1-line-diff-ckpt-60), the model is designed to generate code with an emphasis on security.
  • Parameter Configuration: The steering process used a theta value of 2.0 and keep_sft set to False, indicating a direct application of the steering vectors without retaining the original SFT (Supervised Fine-Tuning) characteristics of the adapters.

Good For

  • Secure Code Development: Ideal for developers and organizations focused on generating code that adheres to security best practices.
  • Research in Model Steering: Provides a practical example of applying task vector arithmetic to modify model behavior for specific outcomes, particularly in the domain of code security.
  • Experimentation with Qwen2.5-Coder-7B: Offers a specialized version of the Qwen2.5-Coder-7B-Instruct base model for tasks requiring a security-conscious coding style.