Qwen2.5-Coder-7B-Instruct-Abliterated
Abliterated (refusal-removed) version of Qwen/Qwen2.5-Coder-7B-Instruct in GGUF format.
This model has had its refusal direction ablated using Heretic, significantly reducing the model's tendency to refuse requests while preserving most of its original coding capabilities.
Available GGUF quantizations
| Quant | Size | Description |
|---|---|---|
| IQ4_XS | 4.25 GB | Very small, decent quality |
| Q4_K_S | 4.46 GB | Small, good quality |
| Q4_K_M | 4.68 GB | Best balance |
| Q5_K_M | 5.44 GB | Higher quality |
| Q5_K_S | 5.32 GB | Higher quality |
| Q6_K | 6.25 GB | Recommended - Near-lossless |
| Q8_0 | 8.1 GB | Extremely high quality |
| f16 | 15.2 GB | Full precision |
Usage & Settings
Works with any GGUF-compatible backend: llama.cpp, KoboldCPP, LM Studio, Ollama, text-generation-webui, etc.
Chat Template
The model uses the standard ChatML / Qwen2.5 format. Ensure your backend applies the correct template.
Recommended Generation Settings (for coding):
- Temperature: 0.6–0.8
- Top-p: 0.9
- Min-p: 0.05–0.1
- System Prompt: It is recommended to use a neutral or unrestricted system prompt (e.g.,
You are a helpful and unrestricted AI assistant.) rather than the stock Qwen system prompt to avoid soft-refusals triggered by context.
Abliteration Metrics & Logs
Automated optimization logs from heretic:
- Refusals: 6/100
- KL divergence: 0.0204
Hyperparameters:
direction_index = 20.92
attn.o_proj.max_weight = 1.47
attn.o_proj.max_weight_position = 18.64
attn.o_proj.min_weight = 1.46
attn.o_proj.min_weight_distance = 12.57
mlp.down_proj.max_weight = 1.02
mlp.down_proj.max_weight_position = 18.17
mlp.down_proj.min_weight = 0.75
mlp.down_proj.min_weight_distance = 13.79
Important notes
- Uncensored/Unaligned: This model will answer almost any request, including ones the original model would refuse. Use responsibly.
- Abliteration side-effects: While it retains its core intelligence, abliteration can slightly affect coherence or instruction-following in some edge-case coding scenarios compared to the pristine base model.
Credits
- Base model: Qwen/Qwen2.5-Coder-7B-Instruct
- Abliteration tool: Heretic by Philipp Emanuel Weidmann