Qwen3-30B-A3B-TopK4-Compressed

이 λͺ¨λΈμ€ Qwen3-30B-A3B Mixture of Experts λͺ¨λΈμ˜ Top-kλ₯Ό 8μ—μ„œ 4둜 κ°μ†Œμ‹œμΌœ μ••μΆ•ν•œ λ²„μ „μž…λ‹ˆλ‹€.

λͺ¨λΈ 정보

  • κΈ°λ³Έ λͺ¨λΈ: Qwen3-30B-A3B
  • μ••μΆ• 방식: Top-k Reduction (8 β†’ 4)
  • μ••μΆ•λ₯ : 59.5%
  • MMLU 정확도: 42.9% (7개 μΉ΄ν…Œκ³ λ¦¬ 평균)

μ••μΆ• 세뢀사항

  • 원본 Top-k: 8개 μ „λ¬Έκ°€ ν™œμ„±ν™”
  • μ••μΆ• Top-k: 4개 μ „λ¬Έκ°€ ν™œμ„±ν™”
  • μ••μΆ• 효과: λͺ¨λΈ 크기와 μΆ”λ‘  λΉ„μš©μ„ 크게 μ€„μ΄λ©΄μ„œλ„ 합리적인 μ„±λŠ₯ μœ μ§€

μ„±λŠ₯ 평가

MMLU 벀치마크 7개 μΉ΄ν…Œκ³ λ¦¬μ—μ„œ ν…ŒμŠ€νŠΈ:

  • abstract_algebra
  • anatomy
  • high_school_mathematics
  • formal_logic
  • professional_medicine
  • high_school_macroeconomics
  • global_facts

μ‚¬μš©λ²•

from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "kyne0127/Qwen3-30B-A3B-TopK4-Compressed"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=True)

# μΆ”λ‘  μ˜ˆμ‹œ
input_text = "What is the capital of France?"
inputs = tokenizer(input_text, return_tensors="pt")
outputs = model.generate(**inputs, max_length=100)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

λΌμ΄μ„ μŠ€

Apache 2.0 License

μ œμž‘μž

kyne0127

Downloads last month
18
Safetensors
Model size
31B params
Tensor type
BF16
Β·
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for kyne0127/Qwen3-30B-A3B-TopK4-Compressed

Finetuned
(92)
this model
Quantizations
2 models