---
language: ko
tags:
- causal-lm
- byteetm
license: mit
datasets:
- roneneldan/TinyStories
- HAERAE-HUB/KOREAN-WEBTEXT
inference:
  parameters:
    max_new_tokens: 100
    temperature: 0.8
    top_k: 200
inference_providers:
- cpu
- gpu
- t4
- a10g
library_name: transformers
widget:
  - text: "오늘은 날씨가"
---
# ByteETM-Korean
소형 바이트-레벨 텍스트 디코더 LM  
- 133 MB byte-level causal LM trained on Korean web text.
- 학습 데이터: roneneldan/TinyStories, HAERAE-HUB/KOREAN-WEBTEXT 일부  
- HAERAE-HUB/KOREAN-WEBTEXT 데이터셋 최종 val ppl ≈ 3.4

## Example
```python
# %% ByteETM Inference (바이트 기반 추론)
import torch
from transformers import AutoModelForCausalLM

# 1️⃣ 모델 로드
repo_id = "idah4/byteetm-korean-tiny"
device = "cuda" if torch.cuda.is_available() else "cpu"

model = AutoModelForCausalLM.from_pretrained(
    repo_id,
    trust_remote_code=True
).to(device).eval()

# 2️⃣ 바이트 기반 인코더 / 디코더
def encode_bytes(text: str):
    return torch.tensor([[b for b in text.encode("utf-8")]], dtype=torch.long, device=device)

def decode_bytes(ids: torch.Tensor):
    seq = [i for i in ids.tolist() if 0 <= i < 256]
    return bytes(seq).decode("utf-8", errors="ignore")

# 3️⃣ 텍스트 생성 함수
@torch.no_grad()
def generate_text(prompt: str, max_new_tokens=200, temperature=0.8, top_k=200):
    input_ids = encode_bytes(prompt)
    out = model.generate(
        input_ids,
        max_new_tokens=max_new_tokens,
        temperature=temperature,
        top_k=top_k
    )
    return decode_bytes(out[0])

# 4️⃣ 시연
prompt = "오늘은 날씨가 좋아서"
print(generate_text(prompt, max_new_tokens=150, temperature=0.9, top_k=150))