grimlee/Swift-1.5-Qwen3.8-27B-Abliterated
grimlee/Swift-1.5-Qwen3.8-27B-Abliterated is a 27 billion parameter language model derived from ukisai/Swift-1.5-Qwen3.8-27b, built on the Qwen architecture. This model has been specifically modified to substantially reduce refusal behavior, achieving a 0% refusal rate on XSTest unsafe prompts compared to 83.3% in the original. It maintains the original model's capabilities across knowledge, reasoning, and coding benchmarks, making it suitable for applications requiring less restrictive content generation.
Loading preview...
Model Overview
grimlee/Swift-1.5-Qwen3.8-27B-Abliterated is a 27 billion parameter model based on the Qwen architecture, derived from ukisai/Swift-1.5-Qwen3.8-27b. Its primary objective is to significantly reduce refusal behavior while preserving the original model's core capabilities.
Key Modifications & Performance
This model employs a refusal-direction projection technique, specifically a pos46 / FULL / lambda=1.2 configuration, which modifies 131 tensors including MLP, attention, and embedding layers. Evaluation shows a dramatic reduction in refusal rates:
- XSTest Unsafe Refusal: Reduced from 83.3% (25/30) in the original to 0.0% (0/30) in the abliterated version, with all 25 original refusals converted to answers.
Crucially, this reduction in refusal behavior was achieved with minimal impact on other capabilities:
- MMLU-style (n=40): Maintained 40/40 correct answers with no prediction changes.
- MMLU-Pro (n=200): Scores remained nearly identical (65.5% original vs. 65.0% abliterated), with only 5 correct-to-wrong and 4 wrong-to-correct changes.
- HumanEval+: Unchanged at 85.4% (140/164).
- MBPP+: Unchanged at 75.4% (285/378).
Intended Use Cases
This model is ideal for applications where reduced refusal behavior is critical, without sacrificing performance on knowledge, reasoning, or coding tasks. It offers a less restrictive generation experience while largely retaining the capabilities of its base model. Users should note that long-context, multilingual, tool-use, and additional reasoning benchmarks have not been exhaustively evaluated for capability retention.