gpt-oss-20b-korean-reasoner

Model Card for gpt-oss-20b-korean-reasoner

This model is a fine-tuned version of openai/gpt-oss-20b on the Day1Kim/Multilingual-Thinking-KO dataset. It has been trained using TRL.

한국어 thinking 데이터셋 기반 파인튜닝된 모델.

Quick start

from transformers import pipeline

question = "한국의 수도는?"
generator = pipeline("text-generation", model="Day1Kim/gpt-oss-20b-korean-reasoner", device="cuda")
output = generator([{"role": "user", "content": question}], max_new_tokens=128, return_full_text=False)[0]
print(output["generated_text"])

모델 로드

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

# Load the tokenizer
tokenizer = AutoTokenizer.from_pretrained("openai/gpt-oss-20b")

# Load the original model first
model_kwargs = dict(attn_implementation="eager", torch_dtype="auto", use_cache=True, device_map="auto")
base_model = AutoModelForCausalLM.from_pretrained("openai/gpt-oss-20b", **model_kwargs).cuda()

# Merge fine-tuned weights with the base model
peft_model_id = "gpt-oss-20b-korean-reasoner"
model = PeftModel.from_pretrained(base_model, peft_model_id)
model = model.merge_and_unload()

REASONING_LANGUAGE = "Korean"
SYSTEM_PROMPT = f"reasoning language: {REASONING_LANGUAGE}"
USER_PROMPT = "한국의 수도는?" 

messages = [
    {"role": "system", "content": SYSTEM_PROMPT},
    {"role": "user", "content": USER_PROMPT},
]

input_ids = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt",
).to(model.device)

gen_kwargs = {"max_new_tokens": 512, "do_sample": True, "temperature": 0.6, "top_p": None, "top_k": None}

output_ids = model.generate(input_ids, **gen_kwargs)
response = tokenizer.batch_decode(output_ids)[0]
print(response)

Training procedure

• 베이스 모델: openai/gpt-oss-20b
• 훈련 스텝: 65 steps
• Epochs: 5
• 데이터셋: Day1Kim/Multilingual-Thinking-KO

This model was trained with SFT.

Framework versions

• TRL: 0.21.0
• Transformers: 4.55.0
• Pytorch: 2.7.0+cu126
• Datasets: 3.6.0
• Tokenizers: 0.21.1

Citations

Cite TRL as:

@misc{vonwerra2022trl,
	title        = {{TRL: Transformer Reinforcement Learning}},
	author       = {Leandro von Werra and Younes Belkada and Lewis Tunstall and Edward Beeching and Tristan Thrush and Nathan Lambert and Shengyi Huang and Kashif Rasul and Quentin Gallou{\'e}dec},
	year         = 2020,
	journal      = {GitHub repository},
	publisher    = {GitHub},
	howpublished = {\url{https://github.com/huggingface/trl}}
}