Qwen2.5 Korean Code Review LLM

Unsloth

Overview

This model is a fine-tuned version of unsloth/qwen2.5-coder-14b-instruct-bnb-4bit. It is optimized for Korean-language code reviews and programming education.

The model was trained using ewhk9887/korean_code_reviews_from_github, a dataset consisting of Korean code reviews collected from GitHub. The fine-tuning process was done using Unsloth and Hugging Face's transformers and trl libraries, enabling a 2x faster training process.

๋ชจ๋ธ ๊ฐœ์š”

์ด ๋ชจ๋ธ์€ unsloth/qwen2.5-coder-14b-instruct-bnb-4bit๋ฅผ ํŒŒ์ธํŠœ๋‹ํ•œ ๋ฒ„์ „์œผ๋กœ, ํ•œ๊ตญ์–ด ์ฝ”๋“œ ๋ฆฌ๋ทฐ ๋ฐ ์ฝ”๋”ฉ ํ•™์Šต์„ ์œ„ํ•œ ์ตœ์ ํ™”๋ฅผ ๊ฑฐ์ณค์Šต๋‹ˆ๋‹ค.

GitHub์—์„œ ์ˆ˜์ง‘๋œ ์ฝ”๋“œ ๋ฆฌ๋ทฐ ๋ฐ์ดํ„ฐ์…‹์„ ์‚ฌ์šฉํ•˜์—ฌ ํ•™์Šตํ–ˆ์œผ๋ฉฐ, Unsloth ๋ฐ Hugging Face์˜ transformers, trl ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋ฅผ ํ™œ์šฉํ•˜์—ฌ 2๋ฐฐ ๋น ๋ฅธ ํ•™์Šต์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ–ˆ์Šต๋‹ˆ๋‹ค.

Features / ํŠน์ง•

  • Korean Code Review Support: Designed specifically for analyzing and reviewing code in Korean.

  • Efficient Fine-Tuning: Utilized bnb-4bit quantization and Unsloth for optimized performance.

  • Bilingual Support: Can process both Korean and English inputs.

  • Transformer-based Model: Leverages Qwen2.5's strong coding capabilities.

  • ํ•œ๊ตญ์–ด ์ฝ”๋“œ ๋ฆฌ๋ทฐ ์ตœ์ ํ™”: ์ฝ”๋“œ ๋ฆฌ๋ทฐ๋ฅผ ํ•œ๊ตญ์–ด๋กœ ๋ถ„์„ํ•˜๊ณ  ์ž‘์„ฑํ•˜๋Š” ๋ฐ ์ตœ์ ํ™”๋˜์—ˆ์Šต๋‹ˆ๋‹ค.

  • ํšจ์œจ์ ์ธ ํŒŒ์ธํŠœ๋‹: bnb-4bit ์–‘์žํ™” ๋ฐ Unsloth ๊ธฐ์ˆ ์„ ํ™œ์šฉํ•˜์—ฌ ๋น ๋ฅธ ํ•™์Šต์ด ๊ฐ€๋Šฅํ–ˆ์Šต๋‹ˆ๋‹ค.

  • ํ•œ์˜ ์ง€์›: ํ•œ๊ตญ์–ด์™€ ์˜์–ด ์ž…๋ ฅ์„ ๋ชจ๋‘ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

  • ๊ฐ•๋ ฅํ•œ ํŠธ๋žœ์Šคํฌ๋จธ ๊ธฐ๋ฐ˜: Qwen2.5 ๋ชจ๋ธ์„ ํ™œ์šฉํ•œ ์ฝ”๋“œ ๋ถ„์„ ์„ฑ๋Šฅ.

Usage / ์‚ฌ์šฉ ๋ฐฉ๋ฒ•

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_name = "ewhk9887/qwen2.5-korean-code-review"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float16, device_map="auto")

inputs = tokenizer("์ฝ”๋“œ๋ฅผ ๋ฆฌ๋ทฐํ•ด ์ฃผ์„ธ์š”: def add(a, b): return a + b", return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=100)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Developer / ๊ฐœ๋ฐœ์ž

  • Name: ์€์€์ˆ˜ (Eunsoo Max Eun)
  • License: Apache-2.0

Acknowledgments / ์ฐธ๊ณ  ์ž๋ฃŒ

Downloads last month
4
Safetensors
Model size
14.8B params
Tensor type
FP16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for ewhk9887/Korean-QWEN-Coder2.5-14B

Quantizations
2 models