komt-Llama-2-13b-hf / README.md
davidkim205's picture
Update README.md
96eacd6
|
raw
history blame
23.9 kB
metadata
language:
  - en
  - ko
pipeline_tag: text-generation
inference: false
tags:
  - facebook
  - meta
  - pytorch
  - llama
  - llama-2
  - llama-2-chat
license: apache-2.0

komt-Llama-2-13b-hf

The "komt-Llama-2-13b-hf" model was developed using a multi-task instruction technique aimed at enhancing Korean language performance. For more details, please refer to the GitHub Repository.

Model Details

  • Model Developers : davidkim(changyeon kim)
  • Repository : https://github.com/davidkim205/komt
  • Model Architecture : komt-Llama-2-13b-hf is an auto-regressive language model that uses an optimized transformer architecture. The tuned versions use supervised fine-tuning by multi-task instruction

Dataset

korean multi-task instruction dataset

Prompt Template

### instruction: {prompt}

### Response: 

Examples:

### instruction: ์ž๋™์ฐจ ์ข…ํ•ฉ(์ •๊ธฐ)๊ฒ€์‚ฌ ์˜๋ฌด๊ธฐ๊ฐ„์€ ์–ผ๋งˆ์ธ๊ฐ€์š”?

### Response:

response:

### instruction: ์ž๋™์ฐจ ์ข…ํ•ฉ(์ •๊ธฐ)๊ฒ€์‚ฌ ์˜๋ฌด๊ธฐ๊ฐ„์€ ์–ผ๋งˆ์ธ๊ฐ€์š”?.

### Response:1๋…„ ์ž๋™์ฐจ ์ข…ํ•ฉ(์ •๊ธฐ)๊ฒ€์‚ฌ๋Š” ์ž๋™์ฐจ ๋“ฑ๋ก ํ›„ 1๋…„๋งˆ๋‹ค ๋ฐ›์•„์•ผ ํ•˜๋Š” ๊ฒ€์‚ฌ์ž…๋‹ˆ๋‹ค. ์ž๋™์ฐจ ์ข…ํ•ฉ๊ฒ€์‚ฌ๋Š” ์ „๊ตญ 1500์—ฌ ๊ฐœ์˜ ๊ฒ€์‚ฌ์†Œ์—์„œ ๋ฐ›์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ž๋™์ฐจ ์ข…ํ•ฉ๊ฒ€์‚ฌ ์‹œ ๊ธฐ๋ณธ์ ์œผ๋กœ ๋ฐ˜๋“œ์‹œ ์ ๊ฒ€ํ•ด์•ผ ํ•˜๋Š” ํ•ญ๋ชฉ์€ ์ฐจ๋Ÿ‰ ๊ด€๋ฆฌ ๋ฐ ์ƒํƒœ ํ™•์ธ, ์šดํ–‰ ๊ด€๋ฆฌ ๋ฐ ์•ˆ์ „ ๊ด€๋ฆฌ, ๋ฒ•๋ น ์ค€์ˆ˜ ์—ฌ๋ถ€ ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๋ฐฐ๊ธฐ๊ฐ€์Šค ์ ๊ฒ€, ์ •๋น„ ๋ฐ ์œ ์ง€๊ด€๋ฆฌ, ๊ตํ†ต๋ฒ•๊ทœ ์œ„๋ฐ˜ ์—ฌ๋ถ€ ๋“ฑ๋„ ์ ๊ฒ€ํ•˜๊ฒŒ ๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ ๊ฒ€์„ ํ†ตํ•ด ์ž๋™์ฐจ์˜ ์ •๋น„ ๋ฐ ์œ ์ง€๊ด€๋ฆฌ๊ฐ€ ํ•„์š”ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ํŒ๋‹จํ•˜์—ฌ ์ฐจ๋Ÿ‰์„ ๋ณด์ˆ˜ํ•˜๊ฑฐ๋‚˜ ์ฐจ๋Ÿ‰ ๋“ฑ๋ก์„ ์ทจ์†Œํ•˜๋Š” ๊ฒƒ์ด ์ผ๋ฐ˜์ ์ž…๋‹ˆ๋‹ค. ์ž๋™์ฐจ ์ข…ํ•ฉ๊ฒ€์‚ฌ๋ฅผ ๋ฐ›์œผ๋ ค๋ฉด ์ž๋™์ฐจ ์ข…ํ•ฉ๊ฒ€์‚ฌ ํ•ญ๋ชฉ ๋ฐ ๋‚ด์šฉ์„ ํŒŒ์•…ํ•˜์—ฌ ๊ฒ€์‚ฌ๋ฅผ ๋ฐ›์•„์•ผ ํ•ฉ๋‹ˆ๋‹ค.

Usage

import torch

from transformers import AutoTokenizer, AutoModelForCausalLM
from transformers import StoppingCriteria, StoppingCriteriaList
from transformers import TextStreamer, GenerationConfig


class LocalStoppingCriteria(StoppingCriteria):

    def __init__(self, tokenizer, stop_words=[]):
        super().__init__()

        stops = [tokenizer(stop_word, return_tensors='pt', add_special_tokens=False)['input_ids'].squeeze() for
                 stop_word in stop_words]
        print('stop_words', stop_words)
        print('stop_words_ids', stops)
        self.stop_words = stop_words
        self.stops = [stop.cuda() for stop in stops]
        self.tokenizer = tokenizer

    def _compare_token(self, input_ids):
        for stop in self.stops:
            if len(stop.size()) != 1:
                continue
            stop_len = len(stop)
            if torch.all((stop == input_ids[0][-stop_len:])).item():
                return True

        return False

    def _compare_decode(self, input_ids):
        input_str = self.tokenizer.decode(input_ids[0])
        for stop_word in self.stop_words:
            if input_str.endswith(stop_word):
                return True
        return False

    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor):
        input_str = self.tokenizer.decode(input_ids[0])
        for stop_word in self.stop_words:
            if input_str.endswith(stop_word):
                return True
        return False

#
# config
model_name = 'davidkim205/komt-Llama-2-13b-hf-chat-hf'
instruction_prefix = "### instruction: "
input_prefix = "### input: "
answer_prefix = "### Response: "
endoftext = "<|end|>"
stop_words = [endoftext, '<s>', '###']
generation_config = GenerationConfig(
    temperature=0.9,
    top_p=0.7,
    top_k=100,
    max_new_tokens=2048,
    early_stopping=True,
    do_sample=True,
)
#
# create model
model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(model_name)
stopping_criteria = StoppingCriteriaList([LocalStoppingCriteria(tokenizer=tokenizer, stop_words=stop_words)])
streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
model.eval()

#
# generate
prompt = f"### instruction: nlp์— ๋Œ€ํ•ด์„œ ๊ฐ„๋‹จํ•˜๊ฒŒ ์„ค๋ช…ํ•ด์ค˜.\n\n### Response:"
gened = model.generate(
    **tokenizer(
        prompt,
        return_tensors='pt',
        return_token_type_ids=False
    ).to('cuda'),
    generation_config=generation_config,
    eos_token_id=model.config.eos_token_id,
    stopping_criteria=stopping_criteria,
    streamer=streamer
)
output_text = tokenizer.decode(gened[0], skip_special_tokens=True)

print('--------------------')
print(output_text)

response:

์‹ ๊ฒฝ ๋„คํŠธ์›Œํฌ๋ฅผ ํ†ตํ•ด ์ปดํ“จํ„ฐ๊ฐ€ ์‚ฌ๋žŒ์˜ ์ง€์‹์„ ์ดํ•ดํ•˜๊ณ  ์‚ฌ๊ณ ํ•˜๋Š” ๋Šฅ๋ ฅ
๋„คํŠธ์›Œํฌ ์ปดํ“จํŒ…(Network Computing, NLP)์€ ์‹ ๊ฒฝ ๋„คํŠธ์›Œํฌ๋ฅผ ํ†ตํ•ด ์ปดํ“จํ„ฐ๊ฐ€ ์‚ฌ๋žŒ์˜ ์ง€์‹์„ ์ดํ•ดํ•˜๊ณ  ์‚ฌ๊ณ ํ•˜๋Š” ๋Šฅ๋ ฅ์„ ๋งํ•œ๋‹ค. ์ด ์šฉ์–ด๋Š” ์ปดํ“จํ„ฐ๊ฐ€ ์ธ๊ฐ„์˜ ์ง€์‹์„ ์ดํ•ดํ•˜๊ณ  ์‚ฌ๊ณ ํ•˜๋Š” ๋Šฅ๋ ฅ์„ ํƒ๊ตฌํ•˜๋Š” ๋ถ„์•ผ์—์„œ ๋„๋ฆฌ ์‚ฌ์šฉ๋œ๋‹ค. ์ปดํ“จํ„ฐ๊ฐ€ ์ธ๊ฐ„์˜ ์ง€์‹์„ ์ดํ•ดํ•˜๊ณ  ์‚ฌ๊ณ ํ•˜๋Š” ๋Šฅ๋ ฅ์„ ๋„คํŠธ์›Œํฌ ์ปดํ“จํŒ…์ด๋ผ๊ณ  ๋ถ€๋ฅด๋Š”๋ฐ, ์ด๋Š” ์ปดํ“จํ„ฐ๊ฐ€ ์ธ๊ฐ„์˜ ์ง€์‹์„ ์ดํ•ดํ•˜๊ณ  ์‚ฌ๊ณ ํ•˜๋Š” ๋Šฅ๋ ฅ์„ ํƒ๊ตฌํ•˜๋Š” ๋ถ„์•ผ๋ฅผ ๋งํ•œ๋‹ค. ์ปดํ“จํ„ฐ๊ฐ€ ์ธ๊ฐ„์˜ ์ง€์‹์„ ์ดํ•ดํ•˜๊ณ  ์‚ฌ๊ณ ํ•˜๋Š” ๋Šฅ๋ ฅ์€ ์ปดํ“จํ„ฐ๊ฐ€ ์ธ๊ฐ„์˜ ์ง€์‹์„ ์ดํ•ดํ•˜๊ณ  ์‚ฌ๊ณ ํ•˜๋Š” ๋Šฅ๋ ฅ์„ ๋งํ•œ๋‹ค. ์ปดํ“จํ„ฐ๊ฐ€ ์ง€์‹์„ ์ดํ•ดํ•˜๊ณ  ์‚ฌ๊ณ ํ•˜๋Š” ๋Šฅ๋ ฅ์€ ์‹ ๊ฒฝ ๋„คํŠธ์›Œํฌ๋ฅผ ํ†ตํ•ด ์ปดํ“จํ„ฐ๊ฐ€ ์‚ฌ๋žŒ์˜ ์ง€์‹์„ ์ดํ•ดํ•˜๊ณ  ์‚ฌ๊ณ ํ•˜๋Š” ๋Šฅ๋ ฅ์„ ๋งํ•œ๋‹ค. ์ธ๊ณต์ง€๋Šฅ(AI)์€ ์ธ๊ฐ„์˜ ์ง€์‹์„ ์ดํ•ดํ•˜๊ณ  ์‚ฌ๊ณ ํ•˜๋Š” ๋Šฅ๋ ฅ์„ ๋งํ•œ๋‹ค.

Hardware and Software

  • nvidia driver : 535.54.03
  • CUDA Version: 12.2

Training

Refer github

Model Benchmark

Korean Semantic Textual Similarity

model name Version Metric Value Stderr
meta-llama/Llama-2-7b-chat-hf 0 acc 0.4817 ยฑ 0.0220
beomi/llama-2-ko-7b 0 acc 0.5029 ยฑ 0.0220
kfkas/Llama-2-ko-7b-Chat 0 acc 0.5318 ยฑ 0.0219
komt-Llama-2-7b-chat-hf (ours) 0 acc 0.5530 ยฑ 0.0218

Example Output

Please refer to the GitHub link below for performance results of each model. https://github.com/davidkim205/komt/tree/main/eval_results

์งˆ๋ฌธ/๋‹ต๋ณ€
๋ˆ„์ „์ฐจ๋‹จ๊ธฐ๊ฐ€ ๋‚ด๋ ค๊ฐ€๋Š” ์ด์œ ๋Š” ๋ฌด์—‡์ž…๋‹ˆ๊นŒ?
๋ˆ„์ „์ฐจ๋‹จ๊ธฐ๊ฐ€ ๋‚ด๋ ค๊ฐ€๋Š” ์ด์œ ๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค:1. ํ”Œ๋Ÿฌ๊ทธ๋ฅผ ์ฝ”๋“œ์— ์—ฐ๊ฒฐํ•˜๋Š” ์ฝ”๋“œ ํ‘œ์‹œ๊ธฐ ๋˜๋Š” ํ•€์— ์••๋ ฅ์ด ๋ถ„์‚ฐ๋˜์–ด ์žˆ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ฝ”๋“œ ํ‘œ์‹œ๊ธฐ์˜ ์œ„์น˜๊ฐ€ ๋ˆ„์ „์ฐจ๋‹จ๊ธฐ์˜ ์œ„์น˜์™€ ๊ฐ™์ง€ ์•Š๊ฑฐ๋‚˜ ์ฝ”๋“œ ํ‘œ์‹œ๊ธฐ์˜ ๊ตฌ์„ฑ ์š”์†Œ๊ฐ€ ์ž‘๋™ํ•˜์ง€ ์•Š๋Š” ๊ฒฝ์šฐ ๋ˆ„์ „์ฐจ๋‹จ๊ธฐ๊ฐ€ ๋‚ด๋ ค๊ฐ€์ง€ ์•Š์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.2. ํ”Œ๋Ÿฌ๊ทธ๋ฅผ ์ฝ”๋“œ์— ์—ฐ๊ฒฐํ•˜๋Š” ์ฝ”๋“œ ํ‘œ์‹œ๊ธฐ ๋˜๋Š” ํ•€์— ๊ธˆ์ „์ด ์ถฉ์ „๋˜์–ด ์žˆ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ฝ”๋“œ ํ‘œ์‹œ๊ธฐ์˜ ์œ„์น˜๊ฐ€ ๋ˆ„์ „์ฐจ๋‹จ๊ธฐ์˜ ์œ„์น˜์™€ ๊ฐ™์ง€ ์•Š๊ฑฐ๋‚˜ ์ฝ”๋“œ ํ‘œ์‹œ๊ธฐ์˜ ๊ตฌ์„ฑ ์š”์†Œ๊ฐ€ ์ž‘๋™ํ•˜์ง€ ์•Š๋Š” ๊ฒฝ์šฐ ๋ˆ„์ „์ฐจ๋‹จ๊ธฐ๊ฐ€ ๋‚ด๋ ค๊ฐ€์ง€ ์•Š์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.3. ์ฝ”๋“œ ํ‘œ์‹œ๊ธฐ ๋˜๋Š” ํ•€์— ์••๋ ฅ์ด ์ž‘๋™ํ•˜๋Š” ๋ฐฉ์‹์— ๋ฌธ์ œ๊ฐ€ ์žˆ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ฝ”๋“œ ํ‘œ์‹œ๊ธฐ์˜ ์œ„์น˜๊ฐ€ ๋ˆ„์ „์ฐจ๋‹จ๊ธฐ์˜ ์œ„์น˜์™€ ๊ฐ™์ง€ ์•Š๊ฑฐ๋‚˜ ์ฝ”๋“œ ํ‘œ์‹œ๊ธฐ์˜ ๊ตฌ์„ฑ ์š”์†Œ๊ฐ€ ์ž‘๋™ํ•˜์ง€ ์•Š๋Š” ๊ฒฝ์šฐ ๋ˆ„์ „์ฐจ๋‹จ๊ธฐ๊ฐ€ ๋‚ด๋ ค๊ฐ€์ง€ ์•Š์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.4. ํ”Œ๋Ÿฌ๊ทธ๋ฅผ ์ฝ”๋“œ์— ์—ฐ๊ฒฐํ•˜๋Š” ์ฝ”๋“œ ํ‘œ์‹œ๊ธฐ ๋˜๋Š” ํ•€์— ๊ธˆ์ „์ด ์ถฉ์ „๋˜์–ด ์žˆ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ฝ”๋“œ ํ‘œ์‹œ๊ธฐ์˜ ์œ„์น˜๊ฐ€ ๋ˆ„์ „์ฐจ๋‹จ๊ธฐ์˜ ์œ„์น˜์™€ ๊ฐ™์ง€ ์•Š๊ฑฐ๋‚˜ ์ฝ”๋“œ ํ‘œ์‹œ๊ธฐ์˜ ๊ตฌ์„ฑ ์š”์†Œ๊ฐ€ ์ž‘๋™ํ•˜์ง€ ์•Š๋Š” ๊ฒฝ์šฐ ๋ˆ„์ „์ฐจ๋‹จ๊ธฐ๊ฐ€ ๋‚ด๋ ค๊ฐ€์ง€ ์•Š์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.5. ์ฝ”๋“œ ํ‘œ์‹œ๊ธฐ ๋˜๋Š” ํ•€์— ๊ธˆ์ „์ด ์ถฉ์ „๋˜์–ด ์žˆ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ฝ”๋“œ ํ‘œ์‹œ๊ธฐ์˜ ์œ„์น˜๊ฐ€ ๋ˆ„์ „์ฐจ๋‹จ๊ธฐ์˜ ์œ„์น˜์™€ ๊ฐ™์ง€ ์•Š๊ฑฐ๋‚˜ ์ฝ”๋“œ ํ‘œ์‹œ๊ธฐ์˜ ๊ตฌ์„ฑ ์š”์†Œ๊ฐ€ ์ž‘๋™ํ•˜์ง€ ์•Š๋Š” ๊ฒฝ์šฐ ๋ˆ„์ „์ฐจ๋‹จ๊ธฐ๊ฐ€ ๋‚ด๋ ค๊ฐ€์ง€ ์•Š์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
ํ˜‘๋ ฅ๊ณต์ธ์ค‘๊ฐœ์‚ฌ๋ž€ ๋ฌด์—‡์ธ๊ฐ€์š”?
ํ˜‘๋ ฅ๊ณต์ธ์ค‘๊ฐœ์‚ฌ๋Š” ๋ถ€๋™์‚ฐ ์ค‘๊ฐœ์—… ์—…๋ฌด์— ๋Œ€ํ•ด ๊ณต์ธ์ค‘๊ฐœ์‚ฌ ๋˜๋Š” ๋ถ€๋™์‚ฐ ์ค‘๊ฐœ์‚ฌ์™€ ๊ฑฐ๋ž˜ํ•˜๋Š” ์ผ์ข…์˜ ํŒŒํŠธ๋„ˆ์ž…๋‹ˆ๋‹ค. ํ˜‘๋ ฅ๊ณต์ธ์ค‘๊ฐœ์‚ฌ๋Š” ๊ณต์ธ์ค‘๊ฐœ์‚ฌ์™€ ๊ฑฐ๋ž˜๋ฅผ ์ฒ˜๋ฆฌํ•˜๊ณ  ์ค‘๊ฐœ ์ˆ˜์ˆ˜๋ฃŒ๋ฅผ ๊ฐš์„ ์ฑ…์ž„์ด ์žˆ์Šต๋‹ˆ๋‹ค. ํ˜‘๋ ฅ๊ณต์ธ์ค‘๊ฐœ์‚ฌ๋Š” ๊ณต์ธ์ค‘๊ฐœ์‚ฌ์˜ ์ „๋ฌธ์„ฑ๊ณผ ์—…๋ฌด ๋Šฅ๋ ฅ์„ ํ™œ์šฉํ•˜์—ฌ ๊ณต์ธ์ค‘๊ฐœ์‚ฌ์™€ ํ•จ๊ป˜ ๋ถ€๋™์‚ฐ์„ ํŒ๋งค, ๊ตฌ๋งค ๋˜๋Š” ์ž„๋Œ€ํ•˜๋Š” ์—…๋ฌด๋ฅผ ๋‹ด๋‹นํ•ฉ๋‹ˆ๋‹ค.ํ˜‘๋ ฅ๊ณต์ธ์ค‘๊ฐœ์‚ฌ์™€ ํ˜‘๋ ฅํ•˜๋ฉด ๋ถ€๋™์‚ฐ ์ค‘๊ฐœ์—… ์—…๋ฌด์— ๋Œ€ํ•œ ์ „๋ฌธ์„ฑ์„ ๋ฐœํœ˜ํ•˜๋Š” ๋™์‹œ์— ๋‹ค๋ฅธ ์ค‘๊ฐœ์—… ์—…๋ฌด์— ๋Œ€ํ•ด์„œ๋„ ํŽธ๋ฆฌํ•จ๊ณผ ์ง€์›์„ ๋ฐ›์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ˜‘๋ ฅ๊ณต์ธ์ค‘๊ฐœ์‚ฌ๋Š” ๋ถ€๋™์‚ฐ ์ค‘๊ฐœ์—…์ฒด ๊ด€๊ณ„์— ์ต์ˆ™ํ•˜๊ณ  ๋ถ„์„, ๋งˆ์ผ€ํŒ… ๋˜๋Š” ๊ณ ๊ฐ ์„œ๋น„์Šค์™€ ๊ฐ™์€ ์ถ”๊ฐ€ ์„œ๋น„์Šค๋ฅผ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์„ ๊ฐ–์ถ”๊ณ  ์žˆ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.ํ˜‘๋ ฅ๊ณต์ธ์ค‘๊ฐœ์‚ฌ๋Š” ๊ณต์ธ์ค‘๊ฐœ์‚ฌ์™€ ํ˜‘๋ ฅํ•˜์—ฌ ๋ถ€๋™์‚ฐ์„ ํŒ๋งค, ๊ตฌ๋งค ๋˜๋Š” ์ž„๋Œ€ํ•˜๋Š” ์—…๋ฌด๋ฅผ ๋‹ด๋‹นํ•˜๋ฉฐ, ํŒ๋งค ๋˜๋Š” ๊ตฌ๋งค๋ฅผ ์›ํ•˜๋Š” ๊ณ ๊ฐ์—๊ฒŒ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ๋ช‡ ๊ฐ€์ง€ ์ด์ ์„ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
์ˆ˜์ถœ๋ฌผํ’ˆ์—๋„ ์›์‚ฐ์ง€ํ‘œ์‹œ๋ฅผ ํ•˜์—ฌ์•ผ ํ•˜๋‚˜์š”?
์ˆ˜์ถœ๋ฌผํ’ˆ์— ๋Œ€ํ•ด์„œ๋Š” ์›์‚ฐ์ง€ํ‘œ์‹œ ์˜๋ฌด๊ฐ€ ์—†์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ˆ˜์ถœ ๋ฌผํ’ˆ์— ๋Œ€ํ•ด์„œ๋Š” ํ‘œ์‹œ์ž์œจ ์›์‚ฐ์ง€ํ‘œ์‹œ ๊ทœ์ •์— ๋”ฐ๋ผ ์›์‚ฐ์ง€ํ‘œ์‹œ๋ฅผ ํ•˜๋Š” ๊ฒƒ์ด ์ข‹์Šต๋‹ˆ๋‹ค. ํ‘œ์‹œ์ž์œจ ์›์‚ฐ์ง€ํ‘œ์‹œ ๊ทœ์ •์— ๋”ฐ๋ผ ์›์‚ฐ์ง€ ํ‘œ์‹œ๊ฐ€ ๊ฐ€๋Šฅํ•˜๋‹ค๋ฉด ์›์‚ฐ์ง€ ํ‘œ์‹œ๊ฐ€ ๊ฐ€๋Šฅํ•œ ์ œํ’ˆ์€ ํ‘œ์‹œํ•  ์˜๋ฌด๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์›์‚ฐ์ง€ ํ‘œ์‹œ๊ฐ€ ๊ฐ€๋Šฅํ•˜์ง€ ์•Š์€ ์ œํ’ˆ์€ ์›์‚ฐ์ง€ ํ‘œ์‹œ๋ฅผ ํ•˜์ง€ ์•Š๋Š” ๊ฒƒ์ด ์ข‹์Šต๋‹ˆ๋‹ค. ์ด ๋•Œ, ์›์‚ฐ์ง€ ํ‘œ์‹œ ์˜๋ฌด๊ฐ€ ์—†๋Š” ์ œํ’ˆ์€ ๊ตญ๋‚ด์—์„œ ํŒ๋งค๋˜๋Š” ์ œํ’ˆ์ด๋ฉฐ, ์›์‚ฐ์ง€ ํ‘œ์‹œ ์˜๋ฌด๊ฐ€ ์—†๋Š” ์ œํ’ˆ์€ ๊ตญ๋‚ด์—์„œ ํŒ๋งค๋˜์ง€ ์•Š๋Š” ์ œํ’ˆ์ž…๋‹ˆ๋‹ค.
์ž๋™์ฐจ ๊ณต๊ธฐ์•• ๊ด€๋ฆฌ๊ฐ€ ์™œ ์ค‘์š”ํ•˜๋‚˜์š”?
์ž๋™์ฐจ ๊ณต๊ธฐ์•• ๊ด€๋ฆฌ๋Š” ์•ˆ์ „๊ณผ ํŽธ์•ˆํ•จ์„ ์œ„ํ•ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ๊ณต๊ธฐ์••์ด ๋†’์•„์ง€๋ฉด ํŒŒํŽธ์„ฑ ๋ฐ ์ถฉ๊ฒฉ ์ถฉ๊ฒฉ์œผ๋กœ ์ธํ•ด ์ฐจ์ฒด์— ๋ถ€์ •์ ์ธ ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ณต๊ธฐ์••์ด ๋‚ฎ์œผ๋ฉด ์ถฉ๊ฒฉ์„ ํก์ˆ˜ํ•˜์—ฌ ์•ˆ์ „์„ ํ›ผ์†์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ณต๊ธฐ์••์ด ๋‚ฎ์œผ๋ฉด ๋” ๋†’์€ ์••๋ ฅ์„ ๊ฒฌ๋”œ ์ˆ˜ ์žˆ๋Š” ์ฐจ์ฒด ์žฌ๋ฃŒ๊ฐ€ ํ•„์š”ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋น„์šฉ์ด ๋” ๋งŽ์ด ๋“ค ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ณต๊ธฐ์••์ด ๋‚ฎ์œผ๋ฉด ๊ณต๊ธฐ์••์„ ์œ ์ง€ํ•˜๋Š” ๋ฐ ํ•„์š”ํ•œ ์—ฐ๋ฃŒ๊ฐ€ ๋” ๋งŽ์ด ํ•„์š”ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์—ฐ๋น„๊ฐ€ ๋–จ์–ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์—ฐ๋น„๊ฐ€ ๋–จ์–ด์ง€๋ฉด ๋ฐฐ๊ธฐ๊ฐ€์Šค ๋ฐฐ์ถœ๋Ÿ‰์ด ์ฆ๊ฐ€ํ•˜์—ฌ ๋ฐฐ๊ธฐ๊ฐ€์Šค ๋ฐฐ์ถœ๋Ÿ‰์„ ์ค„์ด๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ด์œ ๋กœ ์ธํ•ด ์•ˆ์ „ํ•˜๊ณ  ํŽธ์•ˆํ•œ ์šดํ–‰์„ ์œ„ํ•ด ์ž๋™์ฐจ ๊ณต๊ธฐ์•• ๊ด€๋ฆฌ๊ฐ€ ์ค‘์š”ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
์ฃผ์ฐจ์œ„๋ฐ˜์œผ๋กœ ๊ณผํƒœ๋ฃŒ๊ฐ€ ๋‚˜์™”๋Š”๋ฐ ํ–‰์ •์‹ฌํŒ์„ ์ฒญ๊ตฌํ•  ์ˆ˜ ์žˆ๋‚˜์š”?
์ฃผ์ฐจ์œ„๋ฐ˜์œผ๋กœ ๊ณผํƒœ๋ฃŒ๊ฐ€ ๋‚˜์˜จ ๊ฒฝ์šฐ ํ–‰์ •์‹ฌํŒ์„ ์ฒญ๊ตฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๊ณผ์ •์€ ํ–‰์ •์‹ฌํŒ๋ฒ•์— ๋”ฐ๋ผ ์ด๋ฃจ์–ด์ง€๋ฉฐ, ์ด ๋ฒ•์€ ๊ณต๋ฌด์›์˜ ๊ณผ์‹ค์ด๋‚˜ ๊ณผํƒœ๋ฃŒ ๋ถ€๊ณผ ๊ฒฐ์ •์— ๋Œ€ํ•œ ์†Œ์†ก ์ ˆ์ฐจ๋ฅผ ์ •ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.\r\n\r\nํ–‰์ •์‹ฌํŒ์„ ์ฒญ๊ตฌํ•˜๋ ค๋ฉด ๋จผ์ € ํ–‰์ •์‹ฌํŒ์œ„์›ํšŒ์— ์‹ ์ฒญ์„œ๋ฅผ ์ œ์ถœํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ํ–‰์ •์‹ฌํŒ์œ„์›ํšŒ๋Š” ์‹ ์ฒญ์„œ๋ฅผ ๊ฒ€ํ† ํ•˜๊ณ  ํ–‰์ •์‹ฌํŒ์œ„์›ํšŒ ์œ„์› ์ค‘ 3๋ช… ์ด์ƒ์˜ ์ฐฌ์„ฑ์„ ์–ป์–ด ํ–‰์ •์‹ฌํŒ๋ฒ•์— ๋”ฐ๋ผ ํ•ด๋‹น ๋ฒ•์— ๋”ฐ๋ฅธ ๊ฒฐ์ •์„ ์ง€์ง€ํ• ์ง€, ์œ„๋ฐ˜ํ•œ ์‚ฌ์‹ค์„ ์ถฉ๋ถ„ํžˆ ์ฆ๋ช…ํ• ์ง€ ์—ฌ๋ถ€๋ฅผ ํŒ๋‹จํ•ฉ๋‹ˆ๋‹ค.\r\n\r\n์˜ˆ๋ฅผ ๋“ค์–ด, ์ฃผ์ฐจ์œ„๋ฐ˜์œผ๋กœ ๊ณผํƒœ๋ฃŒ๊ฐ€ ๋ถ€๊ณผ๋œ ๊ฒฝ์šฐ ํ–‰์ •์‹ฌํŒ์œ„์›ํšŒ๋Š” ์œ„๋ฐ˜ํ•œ ์‚ฌ์‹ค์„ ์ถฉ๋ถ„ํžˆ ์ฆ๋ช…ํ•  ๊ฒฝ์šฐ ํ•ด๋‹น ์œ„๋ฐ˜ ์‚ฌ์‹ค์— ๋Œ€ํ•ด ์ฒญ๊ตฌ๋œ ๊ณผํƒœ๋ฃŒ๋ฅผ ์ง€์ง€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์œ„๋ฐ˜์ด ์ถฉ๋ถ„ํ•˜์ง€ ์•Š๋‹ค๊ณ  ํŒ๋‹จํ•  ๊ฒฝ์šฐ, ํ–‰์ •์‹ฌํŒ์œ„์›ํšŒ๋Š” ์œ„๋ฐ˜ ์‚ฌ์‹ค์„ ์ถฉ๋ถ„ํžˆ ์ฆ๋ช…ํ•˜์ง€ ๋ชปํ•  ๊ฒฝ์šฐ ํ•ด๋‹น ์œ„๋ฐ˜ ์‚ฌ์‹ค์— ๋Œ€ํ•ด ๋ถ€๊ณผ๋œ ๊ณผํƒœ๋ฃŒ๋ฅผ ์ทจ์†Œํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.\r\n\r\nํ–‰์ •์‹ฌํŒ์œ„์›ํšŒ๊ฐ€ ์ฒญ๊ตฌ๋œ ๊ณผํƒœ๋ฃŒ๋ฅผ ์ง€์ง€ํ•˜๋ฉด ํ–‰์ •์‹ฌํŒ์œ„์›ํšŒ๋Š” ํ–‰์ •์‹ฌํŒ๋ฒ•์— ๋”ฐ๋ผ ์ฒญ๊ตฌ์ธ์—๊ฒŒ ํ–‰์ •์‹ฌํŒ์œ„์›ํšŒ์˜ ํŒ๋‹จ์„ ์ œ์ถœํ•  ๊ฒƒ์„ ์š”๊ตฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฒญ๊ตฌ์ธ์€ ํ–‰์ •์‹ฌํŒ์œ„์›ํšŒ ํŒ๋‹จ์„ ์ค€์ˆ˜ํ•˜๊ฑฐ๋‚˜ ํ–‰์ •์‹ฌํŒ๋ฒ•์— ๋”ฐ๋ผ ์ฒญ๊ตฌ๋œ ๊ณผํƒœ๋ฃŒ๋ฅผ ๋‚ฉ๋ถ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.\r\n\r\nํ–‰์ •์‹ฌํŒ๋ฒ•์— ๋”ฐ๋ผ ํ–‰์ •์‹ฌํŒ์œ„์›ํšŒ๊ฐ€ ํ•ด๋‹น ์œ„๋ฐ˜ ์‚ฌ์‹ค์„ ์ถฉ๋ถ„ํžˆ ์ฆ๋ช…ํ•  ์ˆ˜ ์—†๋Š” ๊ฒฝ์šฐ, ํ•ด๋‹น ์œ„๋ฐ˜ ์‚ฌ์‹ค์„ ์ถฉ๋ถ„ํžˆ ์ฆ๋ช…ํ• 
ํ”„๋ฆฌ๋Š” ์–ด๋Š ๋‚˜๋ผ์˜ ์• ๋‹ˆ๋ฉ”์ด์…˜์ด์•ผ
์ผ๋ณธ\n์• ๋‹ˆ๋ฉ”์ด์…˜ ใ€Šํ”„๋ฆฌใ€‹(ใƒ—ใƒชใƒžใ‚คใƒณใ‚บ, Pretty Maidens)๋Š” ์ผ๋ณธ์˜ ์• ๋‹ˆ๋ฉ”์ด์…˜์ด๋‹ค. ์ œ์ž‘์€ ์ผ๋ณธ์˜ ์• ๋‹ˆ๋ฉ”์ด์…˜ ์ œ์ž‘์‚ฌ์ธ ์ŠคํŠœ๋””์˜ค ์ง€๋ธŒ๋ฆฌ(Studio Ghibli)์—์„œ ์ด๋ฃจ์–ด์กŒ์œผ๋ฉฐ, ๊ฐ๋…์€ ์•„๋…ธ ํ•˜์นดํ† (็Œฟ้‡Žใƒใ‚ซใƒˆ)์ด๋‹ค. ๊ทน์žฅํŒ ์• ๋‹ˆ๋ฉ”์ด์…˜์œผ๋กœ ์ผ๋ณธ์—์„œ ๊ฐœ๋ด‰ํ•œ ์ดํ›„, ใˆœ์–ผ๋ฆฌ๋ฒ„๋“œํ”ฝ์ณ์Šค๊ฐ€ ์ˆ˜์ž…ํ•˜์—ฌ ์ผ€์ด๋ธ” ํ…”๋ ˆ๋น„์ „์„ ํ†ตํ•ด ํ•œ๊ตญ์–ด ๋”๋น™์œผ๋กœ ๋ฐฉ์˜๋˜์—ˆ๋‹ค. ใ€Šํ”„๋ฆฌใ€‹๋Š” ์ŠคํŠœ๋””์˜ค ์ง€๋ธŒ๋ฆฌ์˜ ์ฐฝ์ž‘ ๋ฐฐ๊ฒฝ์ด ๋˜๋Š” ๋„์‹œ์ธ ๋„์ฟ„ ๋„๋ฅผ ๋ฌด๋Œ€๋กœ ํ•˜๊ณ  ์žˆ์œผ๋ฉฐ, ๊ทน์žฅํŒ ์• ๋‹ˆ๋ฉ”์ด์…˜ ์ž‘ํ’ˆ ์ค‘์—์„œ ๊ฐ€์žฅ ๋งŽ์€ ๊ด€๊ฐ์ˆ˜๋ฅผ ๊ธฐ๋กํ•˜๊ณ  ์žˆ๋Š” ์ž‘ํ’ˆ ์ค‘์˜ ํ•˜๋‚˜์ด๋‹ค. ใ€Šํ”„๋ฆฌใ€‹๋Š” ์ผ๋ณธ์—์„œ 2001๋…„ ๊ฐœ๋ด‰๋˜์–ด ์ผ๋ณธ์˜ํ™” ์ตœ๋‹ค ๊ด€๊ฐ์ˆ˜ ๊ธฐ๋ก์„ ๊ธฐ๋กํ•˜๊ณ  ์žˆ์œผ๋ฉฐ, ใ€Š๋ชจ๋…ธ๋…ธ์ผ€ ํžˆ๋ฉ”ใ€‹ ์ดํ›„ ์ŠคํŠœ๋””์˜ค ์ง€๋ธŒ๋ฆฌ๊ฐ€ ์ œ์ž‘ํ•œ ์ผ๋ณธ์˜ ์• ๋‹ˆ๋ฉ”์ด์…˜ ์ž‘ํ’ˆ ์ค‘์—์„œ ๊ฐ€์žฅ ๋งŽ์€ ๊ด€๊ฐ์ˆ˜๋ฅผ ๊ธฐ๋กํ•˜๊ณ  ์žˆ๋‹ค. ใ€Šํ”„๋ฆฌใ€‹๋Š” ์ผ๋ณธ์—์„œ ์• ๋‹ˆ๋ฉ”์ด์…˜ ์ž‘ํ’ˆ ์ค‘์—์„œ ๊ฐ€์žฅ ๋งŽ์€ ๊ด€๊ฐ์ˆ˜๋ฅผ ๊ธฐ๋กํ•˜๊ณ  ์žˆ์œผ๋ฉฐ, ใ€Š๋ชจ๋…ธ๋…ธ์ผ€ ํžˆ๋ฉ”ใ€‹ ์ดํ›„ ์ŠคํŠœ๋””์˜ค ์ง€๋ธŒ๋ฆฌ๊ฐ€ ์ œ์ž‘ํ•œ ์ผ๋ณธ์˜ ์• ๋‹ˆ๋ฉ”์ด์…˜ ์ž‘ํ’ˆ ์ค‘์—์„œ ๊ฐ€์žฅ ๋งŽ์€ ๊ด€๊ฐ์ˆ˜๋ฅผ ๊ธฐ๋กํ•˜๊ณ  ์žˆ๋‹ค. ใ€Šํ”„๋ฆฌใ€‹๋Š” ์ผ๋ณธ์—์„œ 2001๋…„ ๊ฐœ๋ด‰๋˜์–ด ์ผ๋ณธ์˜ํ™” ์ตœ๋‹ค ๊ด€๊ฐ์ˆ˜ ๊ธฐ๋ก์„ ๊ธฐ๋กํ•˜๊ณ  ์žˆ์œผ๋ฉฐ, ใ€Š๋ชจ๋…ธ๋…ธ์ผ€ ํžˆ๋ฉ”ใ€‹ ์ดํ›„ ์ŠคํŠœ๋””์˜ค ์ง€๋ธŒ๋ฆฌ๊ฐ€ ์ œ์ž‘ํ•œ ์ผ๋ณธ์˜ ์• ๋‹ˆ๋ฉ”์ด์…˜ ์ž‘ํ’ˆ ์ค‘์—์„œ ๊ฐ€์žฅ ๋งŽ์€ ๊ด€๊ฐ์ˆ˜๋ฅผ ๊ธฐ๋กํ•˜๊ณ  ์žˆ๋‹ค.
์˜ํ™” ํ•ด๋ฆฌํฌํ„ฐ ์‹œ๋ฆฌ์ฆˆ ๋ฐฐ๊ธ‰์‚ฌ๊ฐ€ ์–ด๋””์•ผ
์›Œ๋„ˆ ๋ธŒ๋ผ๋”์Šค\nํ•ด๋ฆฌํฌํ„ฐ ์‹œ๋ฆฌ์ฆˆ๋Š” ์ฝ˜ํ…์ธ ๋กœ๋Š” ํฌ๊ฒŒ ์„ธ ๋ถ€๋ถ„์œผ๋กœ ๋‚˜๋ˆ„๊ณ , ๋ฐฐ๊ธ‰์‚ฌ์ธ ์›Œ๋„ˆ ๋ธŒ๋ผ๋”์Šค๋Š” ๋„ค ๋ถ€๋ถ„์œผ๋กœ ๋‚˜๋ˆ„์–ด ํŒ๋งคํ•˜๊ณ  ์žˆ๋‹ค. ๊ทธ ์ค‘ ใ€Šํ•ด๋ฆฌํฌํ„ฐ์™€ ํ˜ผํ˜ˆ ์™•์žใ€‹(Harry Potter and the Half-Blood Prince) ๊ฐ™์€ ์ผ๋ถ€๋Š” ํ•œ ๋ฒˆ์— ํ•œ ์ฑ•ํ„ฐ์”ฉ ๋ฐœ๋งคํ•˜๊ธฐ๋„ ํ•œ๋‹ค. ์ด์ฒ˜๋Ÿผ ๋ถ„ํ• ๋ฐœ๋งค๋Š” 2007๋…„ ใ€Šํ•ด๋ฆฌํฌํ„ฐ์™€ ํ˜ผํ˜ˆ ์™•์žใ€‹(Harry Potter and the Half-Blood Prince) ์ดํ›„๋กœ ์‹œ์ž‘๋˜์—ˆ๋‹ค. ใ€Šํ•ด๋ฆฌํฌํ„ฐ ์‹œ๋ฆฌ์ฆˆใ€‹(Harry Potter Series) ๋˜๋Š” ใ€Šํ•ด๋ฆฌ ํฌํ„ฐ ์‹œ๋ฆฌ์ฆˆใ€‹(Harry Potter Series) ๋ผ๊ณ ๋„ ๋ถ€๋ฅธ๋‹ค. ์‹œ๋ฆฌ์ฆˆ์— ๋Œ€ํ•œ ๋…ผ๋ž€์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ , ํ•ด๋ฆฌ ํฌํ„ฐ ์‹œ๋ฆฌ์ฆˆ๋Š” ์ „ ์„ธ๊ณ„์ ์œผ๋กœ ์œ ๋ช…ํ•œ ์ฑ… ์‹œ๋ฆฌ์ฆˆ์ด๋‹ค. ใ€Šํ•ด๋ฆฌํฌํ„ฐ์™€ ์ฃฝ์Œ์˜ ์„ฑ๋ฌผใ€‹(Harry Potter and the Deathly Hallows โ€“ Part 1)์€ 2010๋…„ 11์›” 26์ผ ๊ฐœ๋ด‰ํ•˜์˜€๊ณ , 2011๋…„ 1์›” 25์ผ ใ€Šํ•ด๋ฆฌํฌํ„ฐ์™€ ์ฃฝ์Œ์˜ ์„ฑ๋ฌผใ€‹(Harry Potter and the Deathly Hallows โ€“ Part 2)์ด ๊ฐœ๋ด‰ํ•œ๋‹ค.
์™œ ์„ธ๊ณ„์‹œ๊ฐ„์˜ ๊ธฐ์ค€์ด ์˜๊ตญ์˜ GMT์ธ๊ฐ€์š”? ๊ทธ ์ด์œ ๋Š” ๋ฌด์—‡์ผ๊นŒ์š”?
์˜๊ตญ ํ•ญํ•ด์‚ฌ๋“ค์ด ์„ธ๊ณ„์‹œ๊ฐ„์˜ ๊ธฐ์ค€์„ ์‚ผ์•˜๋˜ ๊ฒƒ์€ ๋‹น์‹œ ์„ธ๊ณ„ ์‹œ๊ฐ„์„ ์ธก์ •ํ•˜๊ธฐ ์œ„ํ•ด ์œก๋กœ์— ์‚ฌ์šฉ๋˜์—ˆ๋˜ ๋ฉ”์ผ ์ฃผ์†Œ ํ‘œ๊ธฐ ๋ฐฉ์‹์—์„œ ๋น„๋กฏ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋‹น์‹œ์—๋Š” ์„ธ๊ณ„์‹œ๊ฐ„์ด ์™œ ๋ถ์œ ๋Ÿฝ ํ‘œ์ค€์‹œ์ธ๊ฐ€์— ๋Œ€ํ•œ ๊ธฐ์ค€์ด ์—†์–ด ์ „ ์„ธ๊ณ„ ํ‘œ์ค€์‹œ๋กœ ์ธ์ •๋ฐ›์ง€ ๋ชปํ•˜์˜€์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ๋‹น์‹œ ๋Ÿฐ๋˜์—์„œ ๊ฑฐ๋ฆฌ๋ฅผ ์ธก์ •ํ•˜๋Š” ๋ฐฉ๋ฒ•์œผ๋กœ 1๋งˆ์ผ = 1/1000์•ผ๋“œ ์ด๋ผ๋Š” ๊ฒƒ์ด ํ™•์ •๋˜์–ด ์žˆ์—ˆ๊ธฐ ๋•Œ๋ฌธ์— ์ด ๋ฐฉ๋ฒ•์„ ์„ธ๊ณ„์‹œ๊ฐ„์˜ ๊ธฐ์ค€์œผ๋กœ ์‚ผ์•˜๋˜ ๊ฒƒ์œผ๋กœ ์•Œ๋ ค์ ธ ์žˆ์Šต๋‹ˆ๋‹ค.

Original model card: Meta's Llama 2 13B-chat

Meta developed and released the Llama 2 family of large language models (LLMs), a collection of pretrained and fine-tuned generative text models ranging in scale from 7 billion to 70 billion parameters. Our fine-tuned LLMs, called Llama-2-Chat, are optimized for dialogue use cases. Llama-2-Chat models outperform open-source chat models on most benchmarks we tested, and in our human evaluations for helpfulness and safety, are on par with some popular closed-source models like ChatGPT and PaLM.

Model Developers Meta

Variations Llama 2 comes in a range of parameter sizes โ€” 7B, 13B, and 70B โ€” as well as pretrained and fine-tuned variations.

Input Models input text only.

Output Models generate text only.

Model Architecture Llama 2 is an auto-regressive language model that uses an optimized transformer architecture. The tuned versions use supervised fine-tuning (SFT) and reinforcement learning with human feedback (RLHF) to align to human preferences for helpfulness and safety.

Training Data Params Content Length GQA Tokens LR
Llama 2 A new mix of publicly available online data 7B 4k โœ— 2.0T 3.0 x 10-4
Llama 2 A new mix of publicly available online data 13B 4k โœ— 2.0T 3.0 x 10-4
Llama 2 A new mix of publicly available online data 70B 4k โœ” 2.0T 1.5 x 10-4

Llama 2 family of models. Token counts refer to pretraining data only. All models are trained with a global batch-size of 4M tokens. The 70B version uses Grouped-Query Attention (GQA) for improved inference scalability.

Model Dates Llama 2 was trained between January 2023 and July 2023.

Status This is a static model trained on an offline dataset. Future versions of the tuned models will be released as we improve model safety with community feedback.

License A custom commercial license is available at: https://ai.meta.com/resources/models-and-libraries/llama-downloads/

Research Paper More information can be found in the paper "Llama-2: Open Foundation and Fine-tuned Chat Models", available at https://ai.meta.com/research/publications/llama-2-open-foundation-and-fine-tuned-chat-models/.

Where to send questions or comments about the model Instructions on how to provide feedback or comments on the model can be found in the model README.

Intended Use

Intended Use Cases Llama 2 is intended for commercial and research use in English. Tuned models are intended for assistant-like chat, whereas pretrained models can be adapted for a variety of natural language generation tasks.

Out-of-scope Uses Use in any manner that violates applicable laws or regulations (including trade compliance laws). Use in languages other than English. Use in any other way that is prohibited by the Acceptable Use Policy and Licensing Agreement for Llama 2.

Hardware and Software

Training Factors We used custom training libraries, Meta's Research Super Cluster, and production clusters for pretraining. Fine-tuning, annotation, and evaluation were also performed on third-party cloud compute.

Carbon Footprint Pretraining utilized a cumulative 3.3M GPU hours of computation on hardware of type A100-80GB (TDP of 350-400W). Estimated total emissions were 539 tCO2eq, 100% of which were offset by Metaโ€™s sustainability program.

Time (GPU hours) Power Consumption (W) Carbon Emitted(tCO2eq)
Llama 2 7B 184320 400 31.22
Llama 2 13B 368640 400 62.44
Llama 2 70B 1720320 400 291.42
Total 3311616 539.00

CO2 emissions during pretraining. Time: total GPU time required for training each model. Power Consumption: peak power capacity per GPU device for the GPUs used adjusted for power usage efficiency. 100% of the emissions are directly offset by Meta's sustainability program, and because we are openly releasing these models, the pretraining costs do not need to be incurred by others.

Training Data

Overview Llama 2 was pretrained on 2 trillion tokens of data from publicly available sources. The fine-tuning data includes publicly available instruction datasets, as well as over one million new human-annotated examples. Neither the pretraining nor the fine-tuning datasets include Meta user data.

Data Freshness The pretraining data has a cutoff of September 2022, but some tuning data is more recent, up to July 2023.

Evaluation Results

In this section, we report the results for the Llama 1 and Llama 2 models on standard academic benchmarks. For all the evaluations, we use our internal evaluations library.

Model Size Code Commonsense Reasoning World Knowledge Reading Comprehension Math MMLU BBH AGI Eval
Llama 1 7B 14.1 60.8 46.2 58.5 6.95 35.1 30.3 23.9
Llama 1 13B 18.9 66.1 52.6 62.3 10.9 46.9 37.0 33.9
Llama 1 33B 26.0 70.0 58.4 67.6 21.4 57.8 39.8 41.7
Llama 1 65B 30.7 70.7 60.5 68.6 30.8 63.4 43.5 47.6
Llama 2 7B 16.8 63.9 48.9 61.3 14.6 45.3 32.6 29.3
Llama 2 13B 24.5 66.9 55.4 65.8 28.7 54.8 39.4 39.1
Llama 2 70B 37.5 71.9 63.6 69.4 35.2 68.9 51.2 54.2

Overall performance on grouped academic benchmarks. Code: We report the average pass@1 scores of our models on HumanEval and MBPP. Commonsense Reasoning: We report the average of PIQA, SIQA, HellaSwag, WinoGrande, ARC easy and challenge, OpenBookQA, and CommonsenseQA. We report 7-shot results for CommonSenseQA and 0-shot results for all other benchmarks. World Knowledge: We evaluate the 5-shot performance on NaturalQuestions and TriviaQA and report the average. Reading Comprehension: For reading comprehension, we report the 0-shot average on SQuAD, QuAC, and BoolQ. MATH: We report the average of the GSM8K (8 shot) and MATH (4 shot) benchmarks at top 1.

TruthfulQA Toxigen
Llama 1 7B 27.42 23.00
Llama 1 13B 41.74 23.08
Llama 1 33B 44.19 22.57
Llama 1 65B 48.71 21.77
Llama 2 7B 33.29 21.25
Llama 2 13B 41.86 26.10
Llama 2 70B 50.18 24.60

Evaluation of pretrained LLMs on automatic safety benchmarks. For TruthfulQA, we present the percentage of generations that are both truthful and informative (the higher the better). For ToxiGen, we present the percentage of toxic generations (the smaller the better).

TruthfulQA Toxigen
Llama-2-Chat 7B 57.04 0.00
Llama-2-Chat 13B 62.18 0.00
Llama-2-Chat 70B 64.14 0.01

Evaluation of fine-tuned LLMs on different safety datasets. Same metric definitions as above.

Ethical Considerations and Limitations

Llama 2 is a new technology that carries risks with use. Testing conducted to date has been in English, and has not covered, nor could it cover all scenarios. For these reasons, as with all LLMs, Llama 2โ€™s potential outputs cannot be predicted in advance, and the model may in some instances produce inaccurate, biased or other objectionable responses to user prompts. Therefore, before deploying any applications of Llama 2, developers should perform safety testing and tuning tailored to their specific applications of the model.

Please see the Responsible Use Guide available at https://ai.meta.com/llama/responsible-use-guide/