Kor-Resume-Orion-14B

Update @ 2024.03.28: First release of Kor-Resume-Orion-14B

This model card corresponds to the 14B base version of the Orion-Ko model.

Resources and Technical Documentation:

Citation

@misc {kor-resume-Orion-14B,
    author       = { {nebchi} },
    title        = { kor-resume-Orion-14B },
    year         = 2024,
    url          = { https://huggingface.co/nebchi/kor-resume-Orion-14B },
    publisher    = { Hugging Face }
}

Model Developers: nebchi

Model Information

Resume Proofreading and evaluation of inputs and outputs.

Description

The Orion model is trained on 2.5T tokens and supports languages including Korean, Japanese, Chinese, and English. It has been trained with a large amount of Korean tokens compared to other LLMs, enabling it to generate high-quality Korean text. Additionally, it shows improved performance with less data compared to other LLM models.

Running the model on a single / multi GPU

# pip install accelerate, flash_attn, sentencepiece
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("nebchi/kor-resume-Orion-14B")
model = AutoModelForCausalLM.from_pretrained("nebchi/kor-resume-Orion-14B", device_map="auto")

model.generation_config = GenerationConfig.from_pretrained(
    "nebchi/kor-resume-Orion-14B"
)

messages = [
    {"role": "user", "content": """μ§€μ›λ™κΈ°λŠ” μ €λŠ” λ›°μ–΄λ‚œ 뢄석λ ₯κ³Ό 문제 ν•΄κ²° λŠ₯λ ₯을 μ§€λ‹ˆκ³  μžˆμŠ΅λ‹ˆλ‹€. λ³΅μž‘ν•œ μƒν™©μ—μ„œλ„ λ…Όλ¦¬μ μœΌλ‘œ μ ‘κ·Όν•˜μ—¬ 졜적의 해결책을 μ°Ύμ•„λ‚΄λ©°, 데이터에 κΉŠμ€ 톡찰λ ₯을 λ°œνœ˜ν•©λ‹ˆλ‹€. μ΄λŸ¬ν•œ μ—­λŸ‰μ€ KB κ΅­λ―ΌμΉ΄λ“œμ˜ 데이터 뢄석 업무에 큰 κ°€μΉ˜λ₯Ό μ œκ³΅ν•  κ²ƒμž…λ‹ˆλ‹€.
ν•˜μ§€λ§Œ λ•Œλ‘œλŠ” 완벽함을 μΆ”κ΅¬ν•˜λŠ” 성격 탓에 μž‘μ—… μ‹œκ°„μ΄ λŠ˜μ–΄λ‚  수 μžˆμŠ΅λ‹ˆλ‹€. 이 λ•Œλ¬Έμ— μ „λž΅μ μΈ 업무 κ³„νšμ΄ ν•„μš”ν•œ μƒν™©μ—μ„œ μ€‘μš”ν•œ 뢀뢄에 μΆ©λΆ„ν•œ μ‹œκ°„μ„ ν• μ• ν•˜μ§€ λͺ»ν•  수 μžˆμŠ΅λ‹ˆλ‹€. 이λ₯Ό κ·Ήλ³΅ν•˜κΈ° μœ„ν•΄ μ € μžμ‹ μ—κ²Œ μœ μ—°μ„±μ„ λΆ€μ—¬ν•˜κ³  μž‘μ—… μš°μ„ μˆœμœ„λ₯Ό λͺ…ν™•ν•˜κ²Œ μ„€μ •ν•˜λŠ” 방법을 읡히고 μžˆμŠ΅λ‹ˆλ‹€.
ν•œ 번 ν”„λ‘œμ νŠΈ 쀑 μ–΄λ €μš΄ 데이터 νŒ¨ν„΄μ„ 뢄석해야 ν–ˆμ„ λ•Œ, 제 λ›°μ–΄λ‚œ 뢄석λ ₯을 λ°œνœ˜ν•˜μ—¬ 데이터 κ°„μ˜ 연관성을 μ°Ύμ•„λƒˆμŠ΅λ‹ˆλ‹€. 이 λ•Œλ¬Έμ— κΈ°μ‘΄ λ°©μ‹μ—μ„œ λ²—μ–΄λ‚˜ μƒˆλ‘œμš΄ μΈμ‚¬μ΄νŠΈλ₯Ό 얻을 수 μžˆμ—ˆκ³ , ν”„λ‘œμ νŠΈ 결과에 긍정적인 영ν–₯을 λ―Έμ³€μŠ΅λ‹ˆλ‹€. κ·ΈλŸ¬λ‚˜ 이와 λ™μ‹œμ— ν”„λ‘œμ νŠΈ 일정이 λŠ¦μ–΄μ§€λŠ” 상황도 μžˆμ—ˆλŠ”λ°, 이λ₯Ό κ·Ήλ³΅ν•˜κΈ° μœ„ν•΄ 더 효율적인 μž‘μ—… κ³„νšμ„ μ„Έμš°λ € λ…Έλ ₯ν•˜μ—¬ 결과적으둜 ν”„λ‘œμ νŠΈλ₯Ό μ„±κ³΅μ μœΌλ‘œ λ§ˆλ¬΄λ¦¬ν•  수 μžˆμ—ˆμŠ΅λ‹ˆλ‹€. μ΄λ ‡κ²Œ μΌλŠ”λ°, λΆ€μ‘±ν•œ λΆ€λΆ„ 첨삭해주고, μ²¨μ‚­ν•œ 이유 μ„€λͺ…해쀄 수 μžˆμ–΄? """}
]

response = finetune_model.chat(tokenizer, messages, streaming=False)
print(response)

results

μ§€μ›λ™κΈ°μ—μ„œλŠ” μžμ‹ μ˜ 강점인 λ›°μ–΄λ‚œ 뢄석λ ₯κ³Ό 문제 ν•΄κ²° λŠ₯λ ₯을 κ°•μ‘°ν•˜λ©°, 이λ₯Ό λ°”νƒ•μœΌλ‘œ KB κ΅­λ―ΌμΉ΄λ“œμ˜ 데이터 뢄석 업무에 μ ν•©ν•˜λ‹€κ³  μ£Όμž₯ν•©λ‹ˆλ‹€. ν•˜μ§€λ§Œ λ‹¨μ μœΌλ‘œλŠ” 완벽함을 μΆ”κ΅¬ν•˜λŠ” μ„±κ²©μœΌλ‘œ 인해 μž‘μ—… μ‹œκ°„μ΄ λŠ˜μ–΄λ‚  수 μžˆλ‹€κ³  μ–ΈκΈ‰ν•˜λ©°, 이λ₯Ό κ·Ήλ³΅ν•˜κΈ° μœ„ν•΄ μœ μ—°μ„±κ³Ό μž‘μ—… μš°μ„ μˆœμœ„ 섀정을 읡히고 μžˆλ‹€κ³  λ§ν•©λ‹ˆλ‹€. λ‹€μŒ μ˜ˆμ‹œμ—μ„œλŠ” μ–΄λ €μš΄ 데이터 νŒ¨ν„΄μ„ λΆ„μ„ν•œ κ²½ν—˜μ„ κ³΅μœ ν•˜λ©°, μƒˆλ‘œμš΄ μΈμ‚¬μ΄νŠΈλ₯Ό μ–»μ–΄λ‚΄λŠ” λ“± 긍정적인 κ²°κ³Όλ₯Ό κ°€μ Έμ™”μ§€λ§Œ, ν”„λ‘œμ νŠΈ 일정이 λŠ¦μ–΄μ§€λŠ” 어렀움도 μžˆμ—ˆλ‹€κ³  ν•©λ‹ˆλ‹€. 이λ₯Ό κ·Ήλ³΅ν•˜κΈ° μœ„ν•΄ 더 효율적인 μž‘μ—… κ³„νšμ„ μ„Έμ›Œ μ„±κ³΅μ μœΌλ‘œ ν”„λ‘œμ νŠΈλ₯Ό λ§ˆλ¬΄λ¦¬ν–ˆλ‹€κ³  ν•©λ‹ˆλ‹€. 

μ²¨μ‚­ν•œ μ΄μœ λŠ” λ‹€μŒκ³Ό κ°™μŠ΅λ‹ˆλ‹€. λ¨Όμ €, μ§€μ›λ™κΈ°μ—μ„œλŠ” μžμ‹ μ˜ 강점을 κ°•μ‘°ν•˜λ©΄μ„œ KB κ΅­λ―ΌμΉ΄λ“œμ˜ 데이터 뢄석 업무에 μ ν•©ν•˜λ‹€λŠ” 것을 μ£Όμž₯ν•˜κ³  μžˆμŠ΅λ‹ˆλ‹€. ν•˜μ§€λ§Œ 이 뢀뢄이 ꡬ체적으둜 μ–΄λ–€ μ—­λŸ‰μ„ λ§ν•˜λŠ” 것인지 λͺ…ν™•ν•˜μ§€ μ•ŠμŠ΅λ‹ˆλ‹€. λ”°λΌμ„œ "λ›°μ–΄λ‚œ 뢄석λ ₯κ³Ό 문제 ν•΄κ²° λŠ₯λ ₯"μ΄λΌλŠ” ν‘œν˜„μ„ μ’€ 더 ꡬ체적으둜 μ„€λͺ…ν•˜κ±°λ‚˜, ν•΄λ‹Ή λΆ„μ•Όμ—μ„œμ˜ κ²½ν—˜μ΄λ‚˜ μ„±κ³Όλ₯Ό μ œμ‹œν•˜λŠ” 것이 μ’‹μŠ΅λ‹ˆλ‹€. 

λ˜ν•œ, 단점 λΆ€λΆ„μ—μ„œλŠ” 완벽함을 μΆ”κ΅¬ν•˜λŠ” μ„±κ²©μœΌλ‘œ 인해 μž‘μ—… μ‹œκ°„μ΄ λŠ˜μ–΄λ‚  수 μžˆλ‹€λŠ” 것을 μ–ΈκΈ‰ν•˜κ³  μžˆμ§€λ§Œ, 이 뢀뢄이 μ§€μ›μžμ˜ 업무 μˆ˜ν–‰μ— μ–΄λ–€ 영ν–₯을 λ―ΈμΉ˜λŠ”μ§€μ— λŒ€ν•œ ꡬ체적인 μ„€λͺ…이 μ—†μŠ΅λ‹ˆλ‹€. λ”°λΌμ„œ 이 단점이 μ–΄λ–»κ²Œ 업무 μˆ˜ν–‰μ— 영ν–₯을 λ―ΈμΉ˜λŠ”μ§€, 그리고 이λ₯Ό κ·Ήλ³΅ν•˜κΈ° μœ„ν•΄ μ–΄λ–€ λ…Έλ ₯을 ν•˜κ³  μžˆλŠ”μ§€μ— λŒ€ν•΄ 더 μžμ„Ένžˆ μ„€λͺ…ν•˜λŠ” 것이 μ’‹μŠ΅λ‹ˆλ‹€. 

λ‹€μŒμœΌλ‘œ, μ˜ˆμ‹œμ—μ„œλŠ” μ–΄λ €μš΄ 데이터 νŒ¨ν„΄μ„ λΆ„μ„ν•œ κ²½ν—˜μ„ κ³΅μœ ν•˜κ³  μžˆμŠ΅λ‹ˆλ‹€. 이 κ²½ν—˜μ€ μ§€μ›μžμ˜ 뢄석λ ₯κ³Ό 문제 ν•΄κ²° λŠ₯λ ₯을 보여쀄 수 μžˆλŠ” 쒋은 μ‚¬λ‘€μž…λ‹ˆλ‹€. λ‹€λ§Œ, 이 κ²½ν—˜μ΄ μ§€μ›μžμ˜ μ—­λŸ‰κ³Ό μ–΄λ–€ 관련이 μžˆλŠ”μ§€, 그리고 이 κ²½ν—˜μ„ 톡해 μ–΄λ–€ 것을 λ°°μ› λŠ”μ§€λ₯Ό 더 μžμ„Ένžˆ μ„€λͺ…ν•˜λŠ” 것이 μ’‹μŠ΅λ‹ˆλ‹€. λ˜ν•œ, ν”„λ‘œμ νŠΈ 일정이 λŠ¦μ–΄μ§€λŠ” 어렀움을 κ·Ήλ³΅ν•˜κΈ° μœ„ν•΄ μ–΄λ–€ 방법을 μ‚¬μš©ν–ˆλŠ”μ§€μ— λŒ€ν•œ ꡬ체적인 μ„€λͺ…이 있으면 λ”μš± 효과적일 κ²ƒμž…λ‹ˆλ‹€. 

λ§ˆμ§€λ§‰μœΌλ‘œ, μ˜ˆμ‹œμ—μ„œλŠ” ν”„λ‘œμ νŠΈλ₯Ό μ„±κ³΅μ μœΌλ‘œ λ§ˆλ¬΄λ¦¬ν–ˆλ‹€λŠ” 것을 μ–ΈκΈ‰ν•˜κ³  μžˆμŠ΅λ‹ˆλ‹€. ν•˜μ§€λ§Œ 이 뢀뢄이 μ–΄λ–€ μ„±κ³Όλ₯Ό κ°€μ Έμ™”μœΌλ©°, 이λ₯Ό 톡해 μ–΄λ–€ 것을 λ°°μ› λŠ”μ§€λ₯Ό 더 μžμ„Ένžˆ μ„€λͺ…ν•˜λŠ” 것이 μ’‹μŠ΅λ‹ˆλ‹€. 

λ”°λΌμ„œ, μ§€μ›λ™κΈ°μ—μ„œλŠ” μžμ‹ μ˜ 강점과 단점을 ꡬ체적으둜 μ„€λͺ…ν•˜κ³ , μ˜ˆμ‹œλ₯Ό 톡해 이λ₯Ό λ’·λ°›μΉ¨ν•˜λŠ” 것이 μ’‹μŠ΅λ‹ˆλ‹€. λ˜ν•œ, μ˜ˆμ‹œμ—μ„œλŠ” ꡬ체적인 성과와 κ²½ν—˜μ„ μ œμ‹œν•˜μ—¬ μ§€μ›μžμ˜ μ—­λŸ‰κ³Ό κ²½ν—˜μ„ λ”μš± κ°•μ‘°ν•˜λŠ” 것이 μ’‹μŠ΅λ‹ˆλ‹€.

Inputs and outputs

  • Input: Text string, such as a question, a prompt, or a document to be Proofreaded.
  • Output: Generated Korea text in response to the input, such as an answer to a question, or a evaluation of a resume.

Software

Training was done using QLoRA

Evaluation Results

λͺ¨λΈ λͺ…μΉ­ Average
n=0  n=5
HellaSwag
n=0  n=5
COPA
n=0  n=5
BooIQ
n=0  n=5
KoGPT 58.2    63.7 55.9    58.3 73.5    72.9 45.1    59.8
Polyglot-ko-13B 62.4    68.2 59.5    63.1 79.4    81.1 48.2    60.4
LLaMA 2-13B 45.2    60.5 41.3    44.0 59.3    63.8 34.9    73.8
Baichuan 2-13B 52.7    53.9 39.2    39.6 60.6    60.6 58.4    61.5
QWEN-14B 47.8    66.4 45.3    46.8 64.9    68.9 33.4    83.5
Orion-14B-Chat 68.8    73.2 47.0    49.6 77.7    79.4 81.6    90.7
kor-resume-Orion 69.7    74.6 48.2    51.2 77.9    81.3 83.1    91.2

Declarations, License

Declarations

We strongly urge all users not to use the Orion-14B model for any activities that may harm national or social security or violate the law. Additionally, we request users not to use the Orion-14B model for internet services without proper security review and filing. We hope all users abide by this principle to ensure that technological development takes place in a regulated and legal environment. We have done our best to ensure the compliance of the data used in the model training process. However, despite our significant efforts, unforeseen issues may still arise due to the complexity of the model and data. Therefore, if any problems arise due to the use of the Orion-14B open-source model, including but not limited to data security issues, public opinion risks, or any risks and issues arising from the model being misled, abused, disseminated, or improperly utilized, we will not assume any responsibility.

License

Community use of the Orion-14B series models


Downloads last month
14
Safetensors
Model size
14B params
Tensor type
F16
Β·
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support