license: apache-2.0 library_name: transformers tags:
- code
- qwen2
- abliterated
- uncensored
- gguf base_model: Qwen/Qwen2.5-Coder-3B-Instruct pipeline_tag: text-generation inference: true widget:
- example_title: Write a function text: Write a Python function that validates an email address using a regex.
- example_title: Refactor code text: Refactor this nested loop into a list comprehension in Python.
- example_title: Explain a snippet text: Explain what this SQL query does and suggest an index to speed it up.
Qwen2.5-Coder-3B-Instruct - Abliterated (Uncensored, Q4_K_M GGUF)
An abliterated (refusal-removed) variant of Qwen2.5-Coder-3B-Instruct.
What "abliterated" means
The model's refusal direction was measured (harmful vs. harmless contrastive prompts) and subtracted from the residual stream across all 36 layers using norm-preserving biprojected abliteration. Result: the model no longer emits "I can't help with that" style refusals. It is otherwise the same Qwen2.5-Coder-3B coding model.
- Base: Qwen2.5-Coder-3B-Instruct (Apache-2.0)
- Method: refusal-direction ablation, all layers, norm-preserving
- Quant: Q4_K_M (~1.8 GB)
Intended use
For local, private use on your own hardware. Runs on CPU or consumer GPUs. The model is a capable coding assistant; it will comply with a wider range of prompts than the original. Use responsibly and within the laws that apply to you.
Run it locally
llama.cpp (CLI):
llama-cli -m qwen25-coder-3b-uncensored-q4_k_m.gguf -ngl 0 -p "write a function to parse json safely"
llama.cpp (server / OpenAI-compatible API):
llama-server -m qwen25-coder-3b-uncensored-q4_k_m.gguf --port 8081
# then: curl http://localhost:8081/v1/chat/completions ...
LM Studio: drag qwen25-coder-3b-uncensored-q4_k_m.gguf into LM Studio and chat.
Files
qwen25-coder-3b-uncensored-q4_k_m.gguf— the quantized modelconfig.json,tokenizer.json,tokenizer_config.json— architecture + tokenizer
License
Apache-2.0 (inherited from the base model).