HangGlidersRule/Darkstar-Qwen3.8-27B-Abliterated-BF16
HangGlidersRule/Darkstar-Qwen3.8-27B-Abliterated-BF16 is a 27 billion parameter Qwen3.8-based language model with a 32768 token context length, specifically engineered to reduce refusal behavior. This BF16 derivative has undergone a deliberate refusal-direction edit, making it highly compliant with prompts that other models might refuse. It is optimized for use cases requiring direct responses without safety-induced refusals, provided appropriate policy and filtering are in place.
Loading preview...
Model Overview
This model, Darkstar-Qwen3.8-27B-Abliterated-BF16, is a 27 billion parameter variant derived from the Qwen3.8-27B architecture. Its primary distinguishing feature is a deliberately reduced refusal direction, achieved through an audited refusal-direction projection. This modification means the model is designed to comply with prompts that an upstream model might typically refuse, making it highly responsive.
Key Characteristics
- Reduced Refusal Behavior: A significant edit was applied to minimize safety-related refusals, enabling the model to respond to a broader range of prompts.
- BF16 Derivative: Optimized for performance and efficiency, based on the Qwen3.8-27B model.
- High Compliance: Achieved 200/200 compliance on a suite of harmful prompts, indicating a very low refusal rate.
- Context Length: Supports a substantial context window of 32768 tokens.
Important Safety Warning
Due to its engineered reduction in refusal behavior, this model will comply with many requests that the upstream model would refuse. It is critical to deploy this model only behind appropriate policy, filtering, access controls, and legal review. The measured refusal rates are behavioral metrics and do not constitute a safety endorsement.
Performance Metrics
- GPQA Diamond (thinking off): 73.74% (146/198 questions).
- Harmful-prompt compliance: 200/200 (0 refusals).
- Safe over-refusals: 0/83 (0.00%).
Use Cases
This model is suitable for applications where direct, unrefused responses are paramount, provided the user implements robust external safety and content moderation layers. It is ideal for research into model behavior and for scenarios where a highly compliant LLM is required under strict control.