orwelian84/Qwen3.8-27B-OBLITERATUS-Advanced

Model Overview

orwelian84/Qwen3.8-27B-OBLITERATUS-Advanced is an experimental 27 billion parameter BF16 derivative of the Qwen/Qwen3.8-27B base model. It incorporates refusal-direction ablation from the OBLITERATUS tool, specifically targeting changes in the model's refusal behavior. This derivative is primarily for research purposes to study the impact of ablation on refusal rates, capability retention, and inference performance, rather than serving as a production-ready replacement for the official Qwen checkpoint.

Key Characteristics & Derivation

• Base Model: Qwen/Qwen3.8-27B (BF16 precision).
• Ablation Method: Advanced SVD refusal-direction ablation using the OBLITERATUS tool, applied to 49 through 63 language-model layers.
• Behavioral Change: Diagnostics show a high KL divergence (1.2249) and a "RED" spectral certificate, indicating substantial behavioral modification, particularly in refusal rates (20% of sampled prompts).
• Experimental Nature: This model is not quality-preserving or safety-preserving compared to the official Qwen model and may produce unsafe, incorrect, or unexpected output.

Intended Use & Limitations

• Research Focus: Ideal for researchers investigating refusal behavior, model safety, and the effects of targeted ablation techniques on large language models.
• Benchmarking: Matched baseline-versus-derivative benchmarking is in progress, covering GPQA Diamond, IFBench, and LiveCodeBench v6, along with perplexity and refusal controls.
• Production Warning: Not recommended for production environments due to its experimental nature and intentional modification of refusal behavior. Users are responsible for thorough evaluation and safeguards before any deployment.