Unexpected results wth XLMR trasformer models

🤗Transformers
1

Hey guys,
I am using xlm-roberta-base with transformers==4.29.2 torch==1.13.1

For any given pair of random text, I always get a high cosine similarity.

model_id = "xlm-roberta-base"
tokenizer = XLMRobertaTokenizerFast.from_pretrained(model_id)
xlmr_model = XLMRobertaModel.from_pretrained(model_id)

xlmr_model.eval()

def get_model_output(text, max_length=30):
    text_input = tokenizer(
    text,
    padding="max_length",
    max_length=max_length,
    truncation=True,
    return_tensors="pt",
    add_special_tokens=True
    )
    print(text_input)
    text_embedding = xlmr_model(
    **text_input
    )

  embedding = text_embedding.pooler_output.flatten().tolist()
  return embedding

def cosine_similarity(v1, v2):
  return float(np.dot(v1, v2) / (np.linalg.norm(v1) * np.linalg.norm(v2)))

text1 = "white joota with blue lace"   # (white shoes with blue lace)
text2 = "vcvbhjook  jjjjj"

embedding1 = get_model_output(text1)
embedding2 = get_model_output(text2)

print(cosine_similarity(embedding1, embedding2)) # 0.99