yejin-korean-1b-v8-sft πŸš€

yejin-korean-1b-v8-sftλŠ” ν•œκ΅­μ–΄ 172GB+ κ³ ν’ˆμ§ˆ μ½”νΌμŠ€λ‘œ 100% λ°”λ‹₯λΆ€ν„° μ‚¬μ „ν•™μŠ΅(Pre-trained from Scratch)된 yejin-korean-1b-v8 베이슀 λͺ¨λΈμ— KMMLU, KoAlpaca, ν•œκ΅­μ–΄ κ΅κ³Όμ„œ μ§€μ‹œλ¬Έμ„ κ²°ν•©ν•˜μ—¬ 지도 λ―Έμ„Έμ‘°μ •(Supervised Fine-Tuning)을 μ™„λ£Œν•œ 1B ν•œκ΅­μ–΄ μ†Œν˜• νŒŒμš΄λ°μ΄μ…˜ λͺ¨λΈμž…λ‹ˆλ‹€.


πŸ“Š λͺ¨λΈ μ•„ν‚€ν…μ²˜ 사양

  • 기반 규격: Llama 3 μ°Έμ‘° μ•„ν‚€ν…μ²˜ (hidden=2048, layers=18, heads=16, kv_heads=4, intermediate=5632, GQA 4:1, RoPE ΞΈ=10000)
  • νŒŒλΌλ―Έν„° 수: μ•½ 9μ–΅ 3,200만 (0.93B)
  • ν† ν¬λ‚˜μ΄μ €: EleutherAI/polyglot-ko-1.3b (μ–΄νœ˜ 크기: 30,003)
  • μ»¨ν…μŠ€νŠΈ 길이: 4,096 토큰
  • 포맷: safetensors (Float32 / BFloat16)

πŸš€ μ‚¬μš© μ˜ˆμ‹œ (Python Transformers)

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "cooler8/yejin-korean-1b-v8-sft"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float32, device_map="auto")

prompt = "### μ§€μ‹œμ‚¬ν•­:\n인곡지λŠ₯μ΄λž€ 무엇인지 μ΄ˆλ“±ν•™μƒλ„ μ•ŒκΈ° μ‰½κ²Œ μ„€λͺ…ν•΄μ€˜.\n\n### λ‹΅λ³€:\n"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

output = model.generate(
    input_ids=inputs["input_ids"],
    attention_mask=inputs["attention_mask"],
    max_new_tokens=200,
    temperature=0.7,
    top_p=0.9,
    repetition_penalty=1.15
)

print(tokenizer.decode(output[0], skip_special_tokens=True))

πŸ’‘ ν”„λ‘¬ν”„νŠΈ 포맷 (Prompt Template)

### μ§€μ‹œμ‚¬ν•­:
{μ‚¬μš©μž 질문 λ˜λŠ” μ§€μ‹œλ¬Έ}

### λ‹΅λ³€:
{λͺ¨λΈ 생성 λ‹΅λ³€}
Downloads last month
36
Safetensors
Model size
0.9B params
Tensor type
F32
Β·
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support