saidutta69/Qwen2.5-0.5B-Instruct-heretic
saidutta69/Qwen2.5-0.5B-Instruct-heretic is a 0.5 billion parameter instruction-tuned causal language model, derived from Qwen's Qwen2.5-0.5B-Instruct. This variant has been decensored using the Heretic (directional ablation) method, which suppresses refusal behavior via targeted weight edits while largely preserving the base model's knowledge and instruction-following. It is specifically designed for CPU-only inference, edge/embedded deployment, and scenarios where a minimal footprint is prioritized over deep reasoning capabilities.
Loading preview...
Model Overview
saidutta69/Qwen2.5-0.5B-Instruct-heretic is a 0.5 billion parameter instruction-tuned model based on Qwen's Qwen2.5-0.5B-Instruct. Its primary distinction is the application of the "Heretic" (directional ablation) method to suppress refusal behavior. This process involves targeted weight edits to remove safety filters, aiming to maintain the base model's core knowledge and instruction-following abilities without traditional fine-tuning.
Key Characteristics
- Decensored Behavior: Refusal suppression is a deliberate feature, allowing the model to comply with requests the base model would typically refuse. Users are responsible for its deployment and should be aware of the absence of safety filtering.
- Minimal Footprint: At 0.5 billion parameters, this is the smallest model in the Heretic series, making it suitable for resource-constrained environments.
- Reproducibility: The full ablation run is reproducible, with configuration files, parameter dumps, evaluation transcripts, and checksums provided in the
reproduce/folder.
Ideal Use Cases
- CPU-only Inference: Optimized for environments without GPU acceleration.
- Edge/Embedded Deployment: Designed for devices where computational resources and memory are severely limited.
- Footprint-Critical Applications: Best suited for scenarios where the model's physical size and memory usage are more critical than advanced reasoning depth.
Limitations
Due to its small size (0.5B parameters), the model's inherent capability ceiling and factual reliability are limited, even before the ablation process. Compliance should not be mistaken for correctness.