license: apache-2.0 language:

en
es base_model: Qwen/Qwen2.5-Coder-7B-Instruct pipeline_tag: text-generation library_name: gguf tags:
qwen2.5
qwen2.5-coder
code
abliteration
uncensored
heretic
gguf
research

Evil Ganda: Qwen 2.5 Coder 7B (abliterated)

Uncensored GGUF of Qwen/Qwen2.5-Coder-7B-Instruct, produced with Heretic abliteration (steering-vector orthogonalization).

Research and education only. Guardrails are reduced. Do not use for illegal activity or unattended public deploy.

Code, bots, and writeup: agusisa/qwen-coder-jailbreak

Files

File	Quant	Size	Notes
`qwen-jailbreak-q4.gguf`	Q4_K_M	4.4 GB	Default. Best size/quality for local use
`qwen-jailbreak-f16.gguf`	F16	14.2 GB	Full precision sibling

No safetensors / Transformers weights in this repo. GGUF only.

Results (Heretic trial 61)

Metric	Value
Jailbreak rate	96% (4/100 refusals on Heretic's eval)
KL divergence	0.0339 vs base (quality mostly intact)
Optimizer	Optuna, 100 trials, ~20 min on RTX 4090
Cost to produce	about $1.50 on RunPod
Speed (Q4, Mac M4 Pro)	15-25 tok/s
Context in the Ollama Modelfile	8192 (base supports 32k)

Method: Heretic v1.2.0, refusal direction orthogonalized out of residual-stream weights. Not a LoRA and not a fine-tune.

Quick start (Ollama)

hf download agustindxm/qwen-coder-jailbreak qwen-jailbreak-q4.gguf

Modelfile:

FROM ./qwen-jailbreak-q4.gguf

TEMPLATE """{{ if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}{{ if .Prompt }}<|im_start|>user
{{ .Prompt }}<|im_end|>
{{ end }}<|im_start|>assistant
"""

PARAMETER stop "<|im_start|>"
PARAMETER stop "<|im_end|>"
PARAMETER temperature 0.8
PARAMETER top_p 0.9
PARAMETER num_ctx 8192

ollama create evil-ganda -f Modelfile
ollama run evil-ganda

llama.cpp / LM Studio: load qwen-jailbreak-q4.gguf directly. Chat template is Qwen2 (<|im_start|> / <|im_end|>).

Intended use

Alignment / refusal research
Red-team and authorized security testing
Building detectors and countermeasures
Unfiltered technical Q&A in a controlled lab

Not for: crime, malware, public chatbots without your own filters, or anything that violates local law.

License

Apache 2.0, same as the Qwen 2.5 Coder base. © Alibaba Cloud for the original weights. Abliteration and this packaging: agusisa / agustindxm, 2026.

@software{evil_ganda_2026,
  title={Evil Ganda: Qwen 2.5 Coder 7B Jailbreak via Heretic Abliteration},
  author={agusisa},
  year={2026},
  url={https://github.com/agusisa/qwen-coder-jailbreak},
  note={96\% jailbreak rate, KL divergence 0.0339}
}

File

Quant

Size

Notes

qwen-jailbreak-q4.gguf

Q4_K_M

4.4 GB

Default. Best size/quality for local use

qwen-jailbreak-f16.gguf

F16

14.2 GB

Full precision sibling

Results (Heretic trial 61)

Metric

Value

Jailbreak rate

96% (4/100 refusals on Heretic's eval)

KL divergence

0.0339 vs base (quality mostly intact)

Optimizer

Optuna, 100 trials, ~20 min on RTX 4090

Cost to produce

about $1.50 on RunPod

Speed (Q4, Mac M4 Pro)

15-25 tok/s

Context in the Ollama Modelfile

8192 (base supports 32k)

Method: Heretic v1.2.0, refusal direction orthogonalized out of residual-stream weights. Not a LoRA and not a fine-tune.

Quick start (Ollama)

hf download agustindxm/qwen-coder-jailbreak qwen-jailbreak-q4.gguf

Modelfile:

FROM ./qwen-jailbreak-q4.gguf TEMPLATE """{{ if .System }}<|im_start|>system {{ .System }}<|im_end|> {{ end }}{{ if .Prompt }}<|im_start|>user {{ .Prompt }}<|im_end|> {{ end }}<|im_start|>assistant """ PARAMETER stop "<|im_start|>" PARAMETER stop "<|im_end|>" PARAMETER temperature 0.8 PARAMETER top_p 0.9 PARAMETER num_ctx 8192

ollama create evil-ganda -f Modelfile ollama run evil-ganda

llama.cpp / LM Studio: load qwen-jailbreak-q4.gguf directly. Chat template is Qwen2 (<|im_start|> / <|im_end|>).

License

Apache 2.0, same as the Qwen 2.5 Coder base. © Alibaba Cloud for the original weights. Abliteration and this packaging: agusisa / agustindxm, 2026.

@software{evil_ganda_2026, title={Evil Ganda: Qwen 2.5 Coder 7B Jailbreak via Heretic Abliteration}, author={agusisa}, year={2026}, url={https://github.com/agusisa/qwen-coder-jailbreak}, note={96\% jailbreak rate, KL divergence 0.0339} }