Modelpublic

Cbgcbg/limo-qwen3-8b-math-full-precision

sourceHugging Faceapache-2.0updated 1y agoView on Hugging Face

0likes8downloads

Model Card

LIMO-Qwen3-8B-Math-Full-Precision

Full-precision (bfloat16) merged model trained with LIMO methodology.

Model Details

—Size: ~16GB (full precision)
—Base: Qwen/Qwen3-8B
—Training: LIMO dataset (817 samples)
—Method: LoRA → Full merge

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "Cbgcbg/limo-qwen3-8b-math-full-precision",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("Cbgcbg/limo-qwen3-8b-math-full-precision")

Comparison

—Original Gasing: 15.26 GB ✅
—Previous LIMO: 5.55 GB ❌ (quantized)
—This model: ~16 GB ✅ (full precision)