felixwangg/Qwen2.5-Coder-7B-steered-alpha-0-variant-A-theta-2.0
The felixwangg/Qwen2.5-Coder-7B-steered-alpha-0-variant-A-theta-2.0 model is a 7.6 billion parameter language model derived from Qwen/Qwen2.5-Coder-7B-Instruct, specifically engineered using task vector arithmetic. This model is steered by applying a scaled difference between 'secure' and 'insecure' adapters, with a theta value of 2.0, to enhance or modify specific behavioral traits. It is designed for applications requiring fine-grained control over model outputs, particularly in code-related tasks, by leveraging its 32768 token context length.
Loading preview...
Model Overview
This model, felixwangg/Qwen2.5-Coder-7B-steered-alpha-0-variant-A-theta-2.0, is a 7.6 billion parameter language model built upon the Qwen/Qwen2.5-Coder-7B-Instruct base. Its unique characteristic lies in its creation via task vector arithmetic, a method that allows for precise steering of model behavior.
Key Capabilities & Steering Mechanism
The model's behavior is determined by the formula: final = pretrained + 2.0 * (TV(secure) - TV(insecure)). This means it combines the base model's capabilities with a weighted difference between a 'secure' adapter (felixwangg/Qwen2.5-Coder-7B-sft-plus-alpha-0-ckpt-30) and an 'insecure' adapter (felixwangg/Qwen2.5-Coder-7B-sft-minus-alpha-0-ckpt-30). The theta parameter, set to 2.0, controls the strength of this steering. This approach enables the model to exhibit specific, desired characteristics by amplifying or diminishing certain learned behaviors.
Use Cases
This model is particularly suited for scenarios where developers need to fine-tune the ethical or safety profile of a code-focused language model without extensive retraining. Its steered nature makes it valuable for:
- Generating code with specific security considerations.
- Research into model steering and behavioral control.
- Developing applications that require nuanced control over model responses based on predefined 'secure' or 'insecure' traits.