jc-lab/multilingual-e5-small-ko-v2-gguf

multilingual-e5-small-ko-v2 GGUF

GGUF quantizations for dragonkue/multilingual-e5-small-ko-v2.

Files:

• ggml-model-f16.gguf
• ggml-model-q8_0.gguf
• ggml-model-q4_k_m.gguf

Verification

The original SentenceTransformer embeddings were compared against normalized embeddings produced by the GGUF models on 7 mixed Korean/English samples.

Summary:

Model	Min cosine to original	Mean cosine to original	Max abs cosine-matrix diff
FP16	0.997235	0.998290	0.012590
Q8_0	0.996689	0.997856	0.014441
Q4_K_M	0.990426	0.992091	0.022216

The full verification output is included in verification_report.json.

Notes

• For Q4_K_M, some tensors fall back to other quantization types because several tensor widths are not divisible by the block-size requirements of Q4_K. The file is still the standard mixed Q4_K_M output produced by llama-quantize.

llama.cpp

./build/bin/llama-embedding \
  -m ggml-model-q8_0.gguf \
  -p "query: 서울에서 맛있는 냉면집 추천해줘"

llama-cpp-python

import numpy as np
import llama_cpp
from llama_cpp import Llama

llm = Llama(
    model_path="ggml-model-q8_0.gguf",
    embedding=True,
    pooling_type=llama_cpp.LLAMA_POOLING_TYPE_MEAN,
    n_ctx=512,
    verbose=False,
)

text = "query: 서울에서 맛있는 냉면집 추천해줘"
vec = np.array(llm.create_embedding(text)["data"][0]["embedding"], dtype=np.float32)
vec = vec / np.linalg.norm(vec)
print(vec.shape)

Use the query: / passage: prefixes exactly as in the original E5 model.