star-bits/gemma-2-2b-lol

Model Card for Model ID

Model Details

Model Description

This model is a fine-tuned version of google/gemma-2-2b-it, designed to answer questions related to champions from the online game League of Legends. By using a custom dataset of champion stories and lore, the model is optimized to generate responses in Korean.

• Developed by: Dohyun Kim, Jongbong Lee, Jaehoon Kim
• Model type: LLM Finetuned Model
• Language(s) (NLP): Korean
• Finetuned from model [optional]: google/gemma-2-2b-it

Training Details

Training Data

The dataset was created by scraping champion lore from the official League of Legends website, transforming the content into Q&A format using large language models. You can find the dataset at fanlino/lol-champion-qa.

# List of champions
champions = [
    "aatrox", "ahri", "akali", "akshan", "alistar", "amumu", "anivia", "annie", "aphelios", "ashe",
    "aurelionsol", "azir", "bard", "belveth", "blitzcrank", "brand", "braum", "caitlyn", "camille",
    "cassiopeia", "chogath", "corki", "darius", "diana", "drmundo", "draven", "ekko", "elise",
    "evelynn", "ezreal", "fiddlesticks", "fiora", "fizz", "galio", "gangplank", "garen", "gnar",
    "gragas", "graves", "gwen", "hecarim", "heimerdinger", "illaoi", "irelia", "ivern", "janna",
    "jarvaniv", "jax", "jayce", "jhin", "jinx", "kaisa", "kalista", "karma", "karthus", "kassadin",
    "katarina", "kayle", "kayn", "kennen", "khazix", "kindred", "kled", "kogmaw", "leblanc", "leesin",
    "leona", "lillia", "lissandra", "lucian", "lulu", "lux", "malphite", "malzahar", "maokai",
    "masteryi", "milio", "missfortune", "mordekaiser", "morgana", "naafiri", "nami", "nasus",
    "nautilus", "neeko", "nidalee", "nilah", "nocturne", "nunu", "olaf", "orianna", "ornn",
    "pantheon", "poppy", "pyke", "qiyana", "quinn", "rakan", "rammus", "reksai", "rell", "renataglasc",
    "renekton", "rengar", "riven", "rumble", "ryze", "samira", "sejuani", "senna", "seraphine", "sett",
    "shaco", "shen", "shyvana", "singed", "sion", "sivir", "skarner", "sona", "soraka", "swain",
    "sylas", "syndra", "tahmkench", "taliyah", "talon", "taric", "teemo", "thresh", "tristana",
    "trundle", "tryndamere", "twistedfate", "twitch", "udyr", "urgot", "varus", "vayne", "veigar",
    "velkoz", "vex", "vi", "viego", "viktor", "vladimir", "volibear", "warwick", "monkeyking", "xayah",
    "xerath", "xinzhao", "yasuo", "yone", "yorick", "yuumi", "zac", "zed", "ziggs", "zilean", "zoe", "zyra"
]

print(f"The total number of champions: {len(champions)}")

# Base URL for the champion story in Korean
base_url = "https://universe.leagueoflegends.com/ko_KR/story/champion/"

# Function to scrape the Korean name and background story of a champion
def scrape_champion_data(champion):
    url = base_url + champion + "/"
    response = requests.get(url)

    if response.status_code == 200:
        soup = BeautifulSoup(response.content, 'html.parser')

        # Extract the Korean name from the <title> tag
        korean_name = soup.find('title').text.split('-')[0].strip()

        # Extract the background story from the meta description
        meta_description = soup.find('meta', {'name': 'description'})
        if meta_description:
            background_story = meta_description.get('content').replace('\n', ' ').strip()
        else:
            background_story = "No background story available"

        return korean_name, background_story
    else:
        return None, None

# Open the CSV file for writing
with open("champion_bs.csv", "w", newline='', encoding='utf-8') as csvfile:
    # Define the column headers
    fieldnames = ['url-name', 'korean-name', 'background-story']

    # Create a CSV writer object
    writer = csv.DictWriter(csvfile, fieldnames=fieldnames)

    # Write the header
    writer.writeheader()

    # Scrape data for each champion and write to CSV
    for champion in champions:
        korean_name, background_story = scrape_champion_data(champion)
        if korean_name and background_story:
            writer.writerow({
                'url-name': champion,
                'korean-name': korean_name,
                'background-story': background_story
            })
            print(f"Scraped data for {champion}: {korean_name}")
        else:
            print(f"Failed to scrape data for {champion}")

print("Data scraping complete. Saved to champion_bs.csv")

Training Procedure

Environment Setup

The model was fine-tuned using a quantization-aware training approach to optimize memory usage and computational efficiency. The environment was set up with 4-bit quantization using torch and transformers, and the LoRA (Low-Rank Adaptation) method was applied to specific layers of the model to improve task performance.

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig

model_id = "google/gemma-2-2b-it"

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_use_double_quant=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16
)

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    quantization_config=qlora_config,
    device_map="auto",
    attn_implementation=attn_implementation
)

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)

QLoRA Setting

from peft import LoraConfig, get_peft_model

def find_linear_layers(model):
    linear_layers = set()
    for name, module in model.named_modules():
        if isinstance(module, bnb.nn.Linear4bit):
            names = name.split('.')
            layer_name = names[-1]
            if layer_name != 'lm_head':
                linear_layers.add(layer_name)
    return list(linear_layers)

lora_target_modules = find_linear_layers(model)

lora_config = LoraConfig(
    r=64,
    lora_alpha=32,
    target_modules=lora_target_modules,
    lora_dropout=0.05,
    bias="none",
    task_type="CAUSAL_LM"
)

model = get_peft_model(model, lora_config)

Loading Training Datasets

To prepare the training data, the champion stories were converted into a question-answer format. The dataset was structured using a chat-style template to ensure compatibility with the Gemma2 model’s architecture.

data = [
{ "q": "대부분의 필멸자가 알고 있는 현실 차원은 무엇인가?", "a": "대부분의 필멸자는 물질 세계라는 하나의 현실 차원만 알고 있다." },
{ "q": "오로라가 유년 시절을 보낸 곳은 어디인가?", "a": "오로라는 브뤼니 부족의 고향이자 외딴 마을인 아무우에서 유년 시절을 보냈다." },
{ "q": "오로라가 자신을 이해해준 유일한 가족 구성원은 누구인가?", "a": "오로라의 이모할머니 하부우가 오로라를 진심으로 받아들였다." },
...]

qa_df = pd.DataFrame(data, columns=["q", "a"])
qa_dataset = Dataset.from_pandas(qa_df)

We use gemma2's chat format template.

<start_of_turn>user
{Qustion}<end_of_turn>
<start_of_turn>model
{Answer}
<end_of_turn>

And we write a function to structure a dataset.

def format_chat_prompt(example):
    chat_data = [
        {"role": "user", "content": example["q"]},
        {"role": "assistant", "content": example["a"]}
    ]
    example["text"] = tokenizer.apply_chat_template(chat_data, tokenize=False)
    return example

dataset = dataset.map(format_chat_prompt, num_proc=4)

The actual format results in the following text.

<bos>
<start_of_turn>user
아트록스가 태어난 곳은 어디인가?<end_of_turn>
<start_of_turn>model
아트록스는 슈리마에서 태어났다.<end_of_turn>'}

Training Model

The model was then trained using the SFTTrainer class, with settings such as a batch size of 1, 10 gradient accumulation steps, and 10 epochs. The optimizer used was paged_adamw_32bit.

import transformers
from trl import SFTTrainer

# Training arguments
training_args = TrainingArguments(
    output_dir=OUTPUT_MODEL_PATH,
    per_device_train_batch_size=1, # steps_per_epoch = ceil(total_samples / (batch_size * gradient_accumulation_steps))
    gradient_accumulation_steps=10, # total_samples means len(dataset)
    num_train_epochs=10,
    learning_rate=2e-4,
    fp16=False,
    bf16=False,
    logging_steps=len(dataset)//10,
    optim="paged_adamw_32bit",
    logging_dir="./logs",
    save_strategy="epoch",
    evaluation_strategy="no",
    do_eval=False,
    group_by_length=True,
    report_to="none"
)

# Initialize trainer
trainer = SFTTrainer(
    model=model,
    train_dataset=dataset,
    peft_config=lora_config,
    dataset_text_field="text",
    max_seq_length=512,
    tokenizer=tokenizer,
    args=training_args,
    packing=False,
)

# Train the model
trainer.train()

Testing Model

We created a helper function to ask the question in the format.

def generate_response(prompt, model, tokenizer, temperature=0.1):
    formatted_prompt=f"""<start_of_turn>user
{prompt}<end_of_turn>
<start_of_turn>model
"""
    inputs = tokenizer(formatted_prompt, return_tensors="pt").to("cuda")
    outputs = model.generate(
        **inputs,
        max_new_tokens=256,
        do_sample=temperature > 0,
        temperature=temperature
    )
    return tokenizer.decode(outputs[0], skip_special_tokens=False)

Question

prompt = "조이는 아우렐리온 솔한테 무슨 약속을 했어?"
response = generate_response(prompt, model, tokenizer)
print(response)

예상 답변

조이는 아우렐리온 솔을 지키기 위해 할 수 있는 것은 무엇이든 해주리라 약속했다.

결과(Finetuned Model)

<bos><start_of_turn>user
조이는 아우렐리온 솔한테 무슨 약속을 했어?<end_of_turn>
<start_of_turn>model
조이는 아우렐리온 솔을 지키기 위해 할 수 있는 것은 무엇이든 해주리라 약속했다.<end_of_turn>

결과(Base Model)

<bos><start_of_turn>user
조이는 아우렐리온 솔한테 무슨 약속을 했어?<end_of_turn>
<start_of_turn>model
 조이는 아우렐리온 솔한테 **무슨 약속을 했는지**에 대한 정보는 아직 알려지지 않았습니다. 

조이는 아우렐리온 솔한테 약속을 했는지에 대한 이야기는 몇 가지 유행하는 밈과 관련된 것으로 보입니다. 

* **아우렐리온 솔:**  이것은 2023년 1월에 출시된 아우렐리온 솔의 이름입니다. 
* **조이:**  이것은 2023년 1월에 출시된 아우렐리온 솔의 이름입니다. 

이러한 밈들은 흥미롭지만, 실제로 조이는 아우렐리온 솔한테 무슨 약속을 했는지에 대한 정확한 정보는 아직 알려지지 않았습니다. 


<end_of_turn>

In contrast, the base model’s response was less accurate, highlighting the improvements made through fine-tuning.

Summary

The code discussed above can be found at the following link: lol_lore.ipynb