felixwangg/Qwen2.5-Coder-7B-steered-alpha-0-variant-A-theta-1.0
The felixwangg/Qwen2.5-Coder-7B-steered-alpha-0-variant-A-theta-1.0 is a 7.6 billion parameter language model, derived from the Qwen2.5-Coder-7B-Instruct base model. This model has been specifically steered using task vector arithmetic to enhance or modify its behavior, combining a 'secure' and 'insecure' adapter. It is designed for specialized applications where fine-grained control over model characteristics, such as security-related responses, is desired. The model leverages a 32K context length for processing extensive code or text inputs.
Loading preview...
Overview
This model, felixwangg/Qwen2.5-Coder-7B-steered-alpha-0-variant-A-theta-1.0, is a 7.6 billion parameter variant of the Qwen/Qwen2.5-Coder-7B-Instruct base model. It has been uniquely modified through task vector arithmetic, a technique that combines different behavioral characteristics into a single model. Specifically, it integrates 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) with a theta value of 1.0.
Key Capabilities
- Behavioral Steering: The model's responses are influenced by a calculated combination of 'secure' and 'insecure' task vectors, allowing for nuanced control over its output characteristics.
- Code-Oriented Base: Built upon the Qwen2.5-Coder-7B-Instruct, it retains strong capabilities in code generation and understanding.
- Customizable Behavior: The steering mechanism offers a method to adjust the model's tendencies without full retraining, making it adaptable for specific safety or ethical considerations.
Good For
- Research in Model Steering: Ideal for exploring the effects of task vector arithmetic on large language models.
- Developing Controlled AI: Useful for scenarios requiring a model whose outputs can be programmatically nudged towards certain attributes (e.g., more secure coding practices or specific response styles).
- Comparative Analysis: Can be used to compare the behavior of a steered model against its base model or other steered variants.