Yejin Korean 8B v3 cpt6000 + Qwen3.6-27B distill r128/lr3e-4 soup6 nothink-v2

Korean instruction-following language model.

Chat Template

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

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("cooler8/yejin-korean-8b-v3-cpt6000-distill-qwen36-r128-lr3e4-soup6-nothink-v2", torch_dtype="bfloat16", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("cooler8/yejin-korean-8b-v3-cpt6000-distill-qwen36-r128-lr3e4-soup6-nothink-v2")

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)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Training

Uniform weight average (model soup) of six LoRA fine-tunes of the same Yejin model: Yejin 8B v3 (pretrained from scratch by Yejin: cooler8/yejin-korean-8b-v3-base) after 6,000 steps of continued pretraining on public Korean data, with the Yejin chat vector from our own SFT model, then LoRA - three runs at rank 128 / lr 2e-4 and three at rank 64 / lr 3e-4, each in a different data order. The LoRA data mixes our phase-1 format data (45,013 samples) with 39,290 short Korean solutions written by Qwen/Qwen3.6-27B (Apache-2.0) for questions from our own puzzle generator, KMMLU train, AI Hub school exams and AI Hub reading comprehension; only solutions that reached the known answer were kept (sequence-level distillation; no teacher weights or logits). No benchmark test items were used, and the teacher solutions share no word 13-gram with KMMLU-Pro, KMMLU test, CLIcK, MuSR or Korean MuSR items.

License

Apache 2.0