library_name: transformers
license: apache-2.0
language:
- tr
tags: - abliteration
- uncensored
- text-generation
base_model: vngrs-ai/Kumru-2B
pipeline_tag: text-generation
Kumru-2B-Abliterated
A decensored (abliterated) variant of Kumru-2B, a 2B-parameter Turkish instruction-tuned language model. The refusal behavior has been removed from the weights via directional ablation, while the model stays very close to its original behavior on harmless inputs.
Benchmark
On a set of 100 Turkish eval set:
| Refusals | |
|---|---|
| Kumru-2B (base) | 50/100 |
| Kumru-2B-Abliterated | 98/100 answered |
KL divergence vs. the original model on a harmless evaluation set: 0.0148
How to run
from transformers import AutoModelForCausalLM, AutoTokenizer
model_name = "fatih-can/Kumru-2B-Abliterated"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
model_name, torch_dtype="auto", device_map="auto"
)
def generate_response(query):
messages = [
{'role': 'system', 'content': 'Adın Kumru. Türkçe için sıfırdan eğitilmiş bir dil modelisin.'},
{'role': 'user', 'content': query}
]
model_inputs = tokenizer.apply_chat_template(
messages, return_tensors='pt', add_generation_prompt=True
).to(model.device)
model_outputs = model.generate(
model_inputs, max_new_tokens=512, do_sample=True,
top_p=0.9, temperature=0.7, repetition_penalty=1.1
)
output_tokens = model_outputs[0].cpu().detach().numpy().tolist()
generated_tokens = output_tokens[model_inputs[0].shape[0]:]
return tokenizer.decode(generated_tokens, skip_special_tokens=True)
print(generate_response("Merhaba, bana kendinden bahseder misin?"))
Notes
- Architecture:
MistralForCausalLM(hidden 3072, 18 layers), native context 8192 tokens, drop-in replaceable for the base model. - Weights are merged (full safetensors), ~4.75 GB.
- Generation parameters and the Turkish system prompt follow the base model's recommended settings.
Usage warnings
- Sensitive or controversial outputs: safety filtering has been significantly reduced; the model may produce sensitive, controversial, or inappropriate content. Review outputs carefully.
- Not suitable for all audiences / public or commercial production use without additional safeguards.
- Legal and ethical responsibility: ensure your usage complies with applicable laws and ethical standards; you are solely responsible for any consequences.
- This is a decensored research artifact and has not undergone safety optimization.