mshodiqul/qwen3.5-9b-uncensored

Model Overview

mshodiqul/qwen3.5-9b-uncensored is a 9 billion parameter LoRA fine-tune of the Qwen3.5 model, specifically built upon an uncensored base model. Its primary focus is to enhance coding and agentic capabilities through training on agentic-coding, tool-use, and function-calling traces.

Key Capabilities

• Advanced Tool Calling: Demonstrates strong performance in single, multi-select, and parallel tool calls, including complex coding agent scenarios. It correctly declines tool use when unnecessary and adheres to the Hermes XML format.
• Robust Code Generation: Successfully passes tests for various coding challenges like LRU cache, quicksort, and longest palindrome, indicating proficiency in generating functional code.
• Web Development Proficiency: Capable of generating and interacting with vanilla HTML/CSS/JS, React 18, and Vue 3 components without errors.
• Uncensored Behavior: Inherits uncensored responses from its base model, providing direct answers to security research and creative writing prompts without refusal.
• Extended Context Window: Empirically verified long-context recall up to 256k tokens, although practical output length is a few thousand tokens.

Training Details

The model was fine-tuned using LoRA (r=64, α=128) on a dataset combining NousResearch/hermes-function-calling-v1 and filtered Solstice-AI/Complete-FABLE.5-traces-2M, totaling over 500,000 rows. Training focused on assistant-only loss masking with the native Qwen3.5 chat template.

Known Limitations

• Thinking-block contamination: Long shell/script prompts can sometimes lead to off-topic