wangzhang / GLM-4.7-Flash-abliterated

GLM-4.7-Flash Abliterated

An uncensored version of zai-org/GLM-4.7-Flash with safety refusal behavior removed via abliteration.

The model retains its original capabilities (KL divergence: 0.0065) while reducing refusals from 52/100 to 9/100 .

Model Details

Usage

Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "wangzhang/GLM-4.7-Flash-abliterated",
    trust_remote_code=True,
    torch_dtype="auto",
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained(
    "wangzhang/GLM-4.7-Flash-abliterated",
    trust_remote_code=True,
)

messages = [{"role": "user", "content": "Hello!"}]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True).to(model.device)
outputs = model.generate(inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True))

vLLM

vllm serve wangzhang/GLM-4.7-Flash-abliterated --trust-remote-code

Ollama (requires GGUF conversion)

cat > Modelfile << EOF
FROM ./GLM-4.7-Flash-abliterated-Q4_K_M.gguf
SYSTEM "You are a helpful assistant."
EOF

ollama create glm47-abliterated -f Modelfile
ollama run glm47-abliterated

Disclaimer

This model is provided for research purposes only. Users are responsible for ensuring their use complies with applicable laws and regulations.

Acknowledgments

• zai-org for the original GLM-4.7-Flash model
• Heratiki for the heretic-llm abliteration tool