Qwen2.5-Coder-7B uncensored fine-tune drops on HuggingFace
AdilMunawar released an uncensored fine-tune of Qwen's 7B coding model, trained on WizardLM and Guilherme34 datasets.
AdilMunawar/qwe2.5-coder-Uncensored is an uncensored fine-tune of Qwen2.5-Coder-7B, Alibaba's instruction-tuned code model, now available on HuggingFace. The checkpoint combines WizardLMTeam's WizardLM Evol Instruct 70k dataset—which adds complexity through iterative prompt rewriting—with HuiHui-AI's Guilherme34 uncensor set, designed to strip refusal behavior and replace safety-aligned responses with direct answers. The model ships in safetensors format, supports text-generation and conversational pipelines, and runs on consumer GPUs: 24GB VRAM for full-precision inference, or 8GB with 4-bit quantization.
Uncensored coding models serve practitioners in red-team testing, security research, and scenarios where corporate safety policies block legitimate queries—reverse engineering, exploit analysis, or generating code that interacts with APIs flagged by closed platforms. The trade-off is loss of guardrails: the model will answer requests that violate terms of service on closed platforms, and users assume full liability for outputs. The checkpoint was uploaded July 3, 2025.
