vwdubb/Swift-Qwen3.8-27B-Uncensored-MTP-Terse-Coder
vwdubb/Swift-Qwen3.8-27B-Uncensored-MTP-Terse-Coder is a 27 billion parameter model derived from Qwen3.8-27B, merging ajgazin's Swift-Qwen3.8-27B-Uncensored-MTP with Shockem's Qwen3.8-27b-Terse-Coder-LoRA. This model features an abliterated safety alignment, making it uncensored, and incorporates a conciseness pass specifically for coding traces. It is designed for research in interpretability, AI-safety, and robustness evaluation, excelling in coding tasks with terse outputs while preserving its vision tower and MTP head for self-speculative decoding and image understanding.
Loading preview...
Swift-Qwen3.8-27B-Uncensored-MTP-Terse-Coder Overview
This model is a 27 billion parameter merge of ajgazin/Swift-Qwen3.8-27B-Uncensored-MTP and Shockem/Qwen3.8-27b-Terse-Coder-LoRA. It builds upon UkisAI's reasoning-efficient Swift-Qwen3.8-27B base, which has undergone a single-direction refusal ablation by orcarouter, effectively removing its safety alignment. The integrated Terse-Coder LoRA adapter, applied via DPO, adds a conciseness pass specifically targeting coding traces.
Key Capabilities
- Uncensored Behavior: Inherits the base's abliteration, complying with requests that standard models would refuse. Intended for legitimate research in AI safety, interpretability, and red-teaming.
- Concise Code Generation: Optimized for terse outputs in coding tasks due to the integrated Terse-Coder adapter.
- Multimodal Support: Preserves the full vision tower and MTP (Multi-Task Prediction) head from the base model, enabling image understanding and self-speculative decoding.
- Research-Oriented: Designed for controlled experiments and evaluation of refusal mechanisms and robustness.
Good For
- AI Safety Research: Studying refusal mechanisms, interpretability, and robustness in large language models.
- Red-Teaming: Evaluating model vulnerabilities and developing stronger safety protocols.
- Code Generation: Generating concise and efficient code, particularly when terse outputs are desired.
- Controlled Experiments: Environments where uncensored model behavior is required for specific research objectives, with appropriate safeguards in place.
Note: This model has not been independently benchmarked for the combined effects of its layered modifications. Users are responsible for implementing their own safety and moderation layers if deploying in production.