TEmA-small

This model is a fine-tuned version of the LaBSE, which is specialized for sentence similarity tasks in Azerbaijan texts. It maps sentences and paragraphs to a 768-dimensional dense vector space, useful for tasks like clustering, semantic search, and more.

Benchmark Results

STSBenchmark	biosses-sts	sickr-sts	sts12-sts	sts13-sts	sts15-sts	sts16-sts	Average Pearson	Model
0.8253	0.7859	0.7924	0.8444	0.7490	0.8141	0.7600	0.7959	TEmA-small
0.7872	0.8303	0.7801	0.7978	0.6963	0.8052	0.7794	0.7823	Cohere/embed-multilingual-v3.0
0.7927	0.6672	0.7758	0.8122	0.7312	0.7831	0.7416	0.7577	BAAI/bge-m3
0.7572	0.8139	0.7328	0.7646	0.6318	0.7542	0.7092	0.7377	intfloat/multilingual-e5-large-instruct
0.7252	0.7801	0.7250	0.6725	0.7446	0.7301	0.7454	0.7318	Cohere/embed-multilingual-v2.0
0.7485	0.7714	0.7271	0.7170	0.6496	0.7570	0.7255	0.7280	intfloat/multilingual-e5-large
0.7245	0.8237	0.6839	0.6570	0.7125	0.7612	0.7386	0.7288	OpenAI/text-embedding-3-large
0.7363	0.8148	0.7067	0.7050	0.6535	0.7514	0.7070	0.7250	sentence-transformers/LaBSE
0.7376	0.7917	0.7190	0.7441	0.6286	0.7461	0.7026	0.7242	intfloat/multilingual-e5-small
0.7192	0.8198	0.7160	0.7338	0.5815	0.7318	0.6973	0.7142	Cohere/embed-multilingual-light-v3.0
0.6960	0.8185	0.6950	0.6752	0.5899	0.7186	0.6790	0.6960	intfloat/multilingual-e5-base
0.5830	0.2486	0.5921	0.5593	0.5559	0.5404	0.5289	0.5155	antoinelouis/colbert-xm

STS-Benchmark

Accuracy Results

• Cosine Distance: 96.63
• Manhattan Distance: 96.52
• Euclidean Distance: 96.57

Usage

from transformers import AutoTokenizer, AutoModel
import torch

# Mean Pooling - Take attention mask into account for correct averaging
def mean_pooling(model_output, attention_mask):
   token_embeddings = model_output[0] #First element of model_output contains all token embeddings
   input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
   return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9)

# Function to normalize embeddings
def normalize_embeddings(embeddings):
   return embeddings / embeddings.norm(dim=1, keepdim=True)

# Sentences we want embeddings for
sentences = [
   "Bu xoşbəxt bir insandır",
   "Bu çox xoşbəxt bir insandır", 
   "Bu gün günəşli bir gündür"
]

# Load model from HuggingFace Hub
tokenizer = AutoTokenizer.from_pretrained('LocalDoc/TEmA-small')
model = AutoModel.from_pretrained('LocalDoc/TEmA-small')

# Tokenize sentences
encoded_input = tokenizer(sentences, padding=True, truncation=True, max_length=128, return_tensors='pt')

# Compute token embeddings
with torch.no_grad():
   model_output = model(**encoded_input)

# Perform pooling
sentence_embeddings = mean_pooling(model_output, encoded_input['attention_mask'])

# Normalize embeddings
sentence_embeddings = normalize_embeddings(sentence_embeddings)

# Calculate cosine similarities
cosine_similarities = torch.nn.functional.cosine_similarity(
   sentence_embeddings[0].unsqueeze(0), 
   sentence_embeddings[1:], 
   dim=1
)

print("Cosine Similarities:")
for i, score in enumerate(cosine_similarities):
   print(f"Sentence 1 <-> Sentence {i+2}: {score:.4f}")