Yejin Korean 8B v3 SFT v2 (no-think)

Korean instruction-following language model.

Same weights as cooler8/yejin-korean-8b-v3-sft-v2. Only the chat template differs: thinking is disabled by default (an empty <think> </think> block is pre-filled after the assistant tag, so the model answers directly). Pass enable_thinking=True to apply_chat_template to restore the original behavior.

Chat Template

<s><|user|>\n{instruction}<|end|>\n<|assistant|>\n<think> </think>{response}<|end|>\n</s>

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("cooler8/yejin-korean-8b-v3-sft-v2-nothink", torch_dtype="bfloat16", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("cooler8/yejin-korean-8b-v3-sft-v2-nothink")

messages = [{"role": "user", "content": "한국의 수도는 어디인가요?"}]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt", return_dict=True).to(model.device)
inputs.pop("token_type_ids", None)
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

# thinking 모드: tokenizer.apply_chat_template(..., enable_thinking=True)

License

Apache 2.0