SmolLM3-3B-Merged-AIJOAH

Models Merged

Subscribing to my YouTube channel AIJOAH and leave comments.

uses only 25% of the fine-tuned model (SmolLM3-3B) and blends it with 75% of the base model (SmolLM3-3B-Base). (della merge) Adds just a bit of instruction-following ability from the fine-tuned model. Keeps the base model’s stability, generalization, and low hallucination rate.

The following models were included in the merge:

• HuggingFaceTB/SmolLM3-3B-Base
• HuggingFaceTB/SmolLM3-3B

Merge Method

This model was merged using the DELLA merge method

Cuda

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "muzerai/SmolLM3-3B-Merged-AIJOAH"
device = "cuda"  # for GPU usage or "cpu" for CPU usage

# load the tokenizer and the model
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
).to(device)

# prepare the model input
prompt = "Give me a brief explanation of gravity in simple terms."
messages_think = [
    {"role": "user", "content": prompt}
]

text = tokenizer.apply_chat_template(
    messages_think,
    tokenize=False,
    add_generation_prompt=True,
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

# Generate the output
generated_ids = model.generate(**model_inputs, max_new_tokens=32768)

# Get and decode the output
output_ids = generated_ids[0][len(model_inputs.input_ids[0]) :]
print(tokenizer.decode(output_ids, skip_special_tokens=True))

MAC

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

# ✅ 모델 경로
model_name = "./SmolLM3-3B-Merged-AIJOAH"

# ✅ Mac M1/M2용 MPS 디바이스 설정
device = torch.device("mps" if torch.backends.mps.is_available() else "cpu")

# ✅ 모델 및 토크나이저 로드
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float16)
model = model.to(device)

# ✅ 프롬프트 입력
prompt = "대한민국 여행일정을 소개해줘"

messages = [
    {
        "role": "system",
        "content": "/think 너는 한국어로만 대답하는 친절한 여행 가이드야. 모든 사고와 응답은 한국어로 해야 해."
    },
    {
        "role": "user",
        "content": prompt
    }
]

# ✅ 템플릿 적용 (chat_template.jinja 기반 자동 구조 생성)
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)

# ✅ 입력 토큰 생성
inputs = tokenizer([text], return_tensors="pt").to(device)

# ✅ 텍스트 생성 (토큰 충분히 설정 + 종료 토큰 명시)
with torch.no_grad():
    generated_ids = model.generate(
        **inputs,
        max_new_tokens=1024,             # 길이 충분히 확보
        temperature=0.7,                 # 생성 다양성
        do_sample=True,                  # 샘플링 활성화
        eos_token_id=tokenizer.eos_token_id  # 종료 토큰 명시
    )

# ✅ 출력 디코딩 (불필요한 시스템/유저 추정 발화 제거)
output = tokenizer.decode(
    generated_ids[0][inputs.input_ids.shape[1]:],
    skip_special_tokens=False
)

# ✅ 다음 질문이 섞이지 않도록 자르기
output = output.split("<|im_start|>user")[0].strip()

# ✅ 최종 결과 출력
print(output)

<think>
Okay, the user is asking for a travel itinerary in Korean. Let me start by understanding the request. They might be planning a trip to Korea and need some travel suggestions or a sample itinerary. Since I need to respond in Korean, I should make sure the content is appropriate and useful.

First, I should consider the types of tours or attractions in Korea. The most popular ones are Seoul, Jeju Island, Busan, and the southern regions like Jeongju and Gyeongju. Maybe I can suggest a 7-day itinerary that covers these areas. Let me outline a typical 7-day plan.

Starting with Seoul, which has a lot of historical sites and modern attractions. Then, moving to Gyeongju for historical sites like the Gwangyusa Cave and the Namsan Park. Jeju Island is perfect for natural beauty and relaxation. Busan is good for seafood and beaches. Jeongju is known for its temples and hanbok performances. Lastly, Gyeongju can be revisited for more history or exploring nearby areas like Andong or Guri.

I should also include a mix of day trips and overnight stays to cover different regions. Maybe suggest a day in Seoul, a day in Gyeongju, a day in Jeju, a day in Busan, and a day in Jeongju. Each city has its own unique experiences. I'll mention popular activities, like visiting the Gwangyusa Cave, Namsan Tower, Jeongja Market, Jeju Volcano, and Busan Port.

I need to make sure the itinerary is balanced between historical, cultural, and natural attractions. Also, include some local cuisine recommendations to enhance the travel experience. Maybe suggest trying hanbok (Korean traditional clothing) at the Bukhansan Museum or in Jeju Island's hanbok shops.

Wait, the user might appreciate some cultural tips, like etiquette in temples or respectful behavior in historical sites. Also, transportation options between cities, such as buses or flights, could be useful for planning. But since they asked for an itinerary, focusing on the activities and places to visit is better.

Let me structure the response with a 7-day plan, each day with a specific city and highlights. Use bullet points for clarity. Also, add a note about the best times to visit each place and some local customs or tips. That should cover their request effectively.
</think>

안녕하세요! 한국 여행일정을 소개해 드리겠습니다. 7일간의 여행일정으로 다양한 문화, 역사, 자연의 아름다움을 체험할 수 있습니다. 아래에서 각 도시의 명소와 활동을 소개해 드리겠습니다. 

**1일: 서울 (Seoul)**
- **명소:** 국립박물관, 종묘, 한양도성, 강남구역
- **활동:** 한양도성 투어 (성곽과 역사관), 한국전쟁기념관, 삼성동 코엑스
- **추천:** 한양궁궄에서 조선시대 생활을 체험해보세요!

**2일: 경주 (Gyeongju)**
- **명소:** 고구려 왕성, 강산성, 무형문화재전수관
- **활동:** 고구려 유적지 탐방, 석굴암
- **추천:** 석굴암의 자연미와 조선시대 미술의 조화를 감상하세요!

**3일: 제주 (Jeju Island)**
- **명소:** 제주백록도, 한라산, 가야사
- **활동:** 한라산 관광, 제주도민문화체험, 노형산 정원
- **추천:** 한라산 해안길을 걸으며 제주의 자연과 역사에 깊이 빠져보세요!

**4일: 부산 (Busan)**
- **명소:** 해운대, 남포동, 태종대
- **활동:** 해운대 해수욕장, 범일시장, 태종대
- **추천:** 해운대 해수욕장에서 해양공원과 휴양림을 함께 즐기세요!

**5일: 청주 (Cheongju)**
- **명소:** 청주 고구려문화공원, 청주시민체육공원
- **활동:** 청주호를 거꾸로 흐르는 철길, 청주산악공원
- **추천:** 청주 시민문화체험장에서 전통문화 체험을 즐기세요!

**6일: 안동 (Andong) & 구리 (Guri)**
- **명소:** 안동백골, 안동민속촌, 구리역사공원
- **활동:** 안동민속촌에서 전통문화 체험, 구리역사공원에 있는 조선시대 거리
-

Limitations

SmolLM3-3B-Merged-AIJOAH can produce text on a variety of topics, but the generated content may not always be factually accurate, logically consistent, or free from biases present in the training data. These models should be used as assistive tools rather than definitive sources of information. Users should always verify important information and critically evaluate any generated content.

Email

[email protected]

License

Apache 2.0