CarrotAI/Rabbit3-Ko-4B

Model Details

• Name: CarrotAI/Rabbit3-Ko-4B
• Version: 4B Instruct
• Base Model: Qwen/Qwen3-4B
• Languages: Korean, English
• Model Type: Large Language Model (Instruction-tuned)

Qwen3-4B 기반의 LLM 모델로 한국어 및 영어 데이터셋을 사용하여 파인튜닝한 한국어 모델입니다.

• 2025.05.16 일반모드로만 사용 가능합니다.

Score

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
gsm8k	3	flexible-extract	5	exact_match	↑	0.8400	±	0.0101
		strict-match	5	exact_match	↑	0.8378	±	0.0102
hrm8k	N/A
- hrm8k_gsm8k	1	none	0	exact_match	↑	0.8196	±	0.0106
- hrm8k_ksm	1	none	0	exact_match	↑	0.0511	±	0.0058
- hrm8k_math	1	none	0	exact_match	↑	0.5539	±	0.0093
- hrm8k_mmmlu	1	none	0	exact_match	↑	0.5362	±	0.0230
- hrm8k_omni_math	1	none	0	exact_match	↑	0.1812	±	0.0088
ifeval	4	none	0	inst_level_loose_acc	↑	0.8753	±	N/A
		none	0	inst_level_strict_acc	↑	0.8609	±	N/A
		none	0	prompt_level_loose_acc	↑	0.8244	±	0.0164
		none	0	prompt_level_strict_acc	↑	0.8078	±	0.0170

Groups	Version	Filter	n-shot	Metric		Value		Stderr
haerae	1	none	0	acc	↑	0.6654	±	0.0140
		none	0	acc_norm	↑	0.6654	±	0.0140
kobest	1	none	0	acc	↑	0.7768	±	0.0057
		none	0	acc_norm	↑	0.5880	±	0.0220
		none	0	f1	↑	0.7764	±	N/A

Groups	Version	Filter	n-shot	Metric		Value		Stderr
kmmlu_direct	2	none	0	exact_match	↑	0.5212	±	0.0026
- kmmlu_direct_applied_science	2	none	0	exact_match	↑	0.4997	±	0.0046
- kmmlu_direct_humss	2	none	0	exact_match	↑	0.5365	±	0.0068
- kmmlu_direct_other	2	none	0	exact_match	↑	0.5130	±	0.0053
- kmmlu_direct_stem	2	none	0	exact_match	↑	0.5455	±	0.0048

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "CarrotAI/Rabbit3-Ko-4B"

# load the tokenizer and the model
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

# prepare the model input
prompt = "Give me a short introduction to large language model."
messages = [
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=True # Switches between thinking and non-thinking modes. Default is True.
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

# conduct text completion
generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=32768
)
output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist() 

# parsing thinking content
try:
    # rindex finding 151668 (</think>)
    index = len(output_ids) - output_ids[::-1].index(151668)
except ValueError:
    index = 0

thinking_content = tokenizer.decode(output_ids[:index], skip_special_tokens=True).strip("\n")
content = tokenizer.decode(output_ids[index:], skip_special_tokens=True).strip("\n")

print("thinking content:", thinking_content)
print("content:", content)

For deployment, you can use sglang>=0.4.6.post1 or vllm>=0.8.5 or to create an OpenAI-compatible API endpoint: