felixwangg/Qwen2.5-Coder-7B-steered-alpha-0-variant-B-theta-1.0
The felixwangg/Qwen2.5-Coder-7B-steered-alpha-0-variant-B-theta-1.0 is a 7.6 billion parameter language model derived from Qwen/Qwen2.5-Coder-7B-Instruct, specifically engineered using task vector arithmetic. This model applies a steering formula to enhance or modify specific behavioral traits, leveraging secure and insecure adapters. It is designed for applications requiring fine-grained control over model behavior, particularly in code-related tasks, by adjusting its responses based on defined steering vectors.
Loading preview...
Model Overview
The felixwangg/Qwen2.5-Coder-7B-steered-alpha-0-variant-B-theta-1.0 is a 7.6 billion parameter language model built upon the Qwen/Qwen2.5-Coder-7B-Instruct base. This model utilizes a technique called Task Vector Arithmetic to steer its behavior, specifically by combining a pretrained model with task vectors derived from "secure" and "insecure" adapters.
Steering Mechanism
The model's final behavior is determined by the formula: final = pretrained + TV(secure) + 1.0 * (TV(secure) - TV(insecure)). This formula indicates that the model's output is influenced by adding the "secure" task vector and further amplifying the difference between the "secure" and "insecure" task vectors by a factor of 1.0. This allows for a directed modification of the base model's characteristics.
Key Components
- Base model:
Qwen/Qwen2.5-Coder-7B-Instruct - Secure adapter:
felixwangg/Qwen2.5-Coder-7B-sft-plus-alpha-0-ckpt-30 - Insecure adapter:
felixwangg/Qwen2.5-Coder-7B-sft-minus-alpha-0-ckpt-30
Parameters
- theta: 1.0 (This parameter controls the strength of the steering effect.)
- keep_sft: True (Indicates that the Supervised Fine-Tuning component is retained.)
Use Cases
This model is particularly suited for scenarios where developers need to:
- Experiment with behavioral steering: Explore how specific traits can be enhanced or suppressed.
- Develop safer code generation: Potentially guide the model towards more secure coding practices by leveraging the "secure" adapter.
- Research model interpretability: Understand the impact of different task vectors on model outputs.