Paralel TR-EN Embedder
A bilingual Turkish-English sentence embedding model fine-tuned on 882K parallel translation pairs. Built on BAAI/bge-small-en-v1.5.
Model Overview
| Property | Value |
|---|---|
| Base Model | BAAI/bge-small-en-v1.5 |
| Embedding Dimension | 384 |
| Max Sequence Length | 64 tokens |
| Parameters | 33M |
| Training Data | 882,555 parallel TR-EN pairs |
| Loss Function | MultipleNegativesRankingLoss |
| GGUF Size | 64 MB (F16) |
Quick Start
Install and load the model:
pip install sentence-transformers
Then in Python:
from sentence_transformers import SentenceTransformer
model = SentenceTransformer("Legenddefsx/paralel-tr-en-embedder")
tr_emb = model.encode("Merhaba, nasilsin?")
en_emb = model.encode("Hello, how are you?")
similarity = model.similarity(tr_emb, en_emb)
GGUF Format
Download and use with llama.cpp:
wget https://huggingface.co/Legenddefsx/paralel-tr-en-embedder/resolve/main/model.gguf
from llama_cpp import Llama
import numpy as np
model = Llama(model_path="./model.gguf", embedding=True)
emb = model.create_embedding("Hello world")["data"][0]["embedding"]
Performance
Evaluated on a held-out set (not seen during training), including correct translations, hard negatives (structurally similar but semantically wrong translations - e.g. wrong tense, wrong pronoun, negation flipped), and unrelated sentence pairs as a control group.
| Category | Avg. Cosine Similarity |
|---|---|
| Correct translations (held-out) | ~0.82 |
| Hard negatives (similar structure, wrong meaning) | ~0.68 |
| Unrelated sentence pairs | ~0.12 |
Known limitation: the model separates unrelated content well, but fine-grained distinctions (pronoun swaps, tense flips, negation) are not yet fully reliable - some hard-negative pairs score close to or above correct-translation pairs. Treat similarity scores as a relevance signal for coarse retrieval/clustering rather than a precise correctness check. Further hard-negative training is in progress.
Use Cases
- Cross-lingual semantic search (coarse retrieval)
- Parallel corpus alignment
- Bilingual document clustering
- Multilingual RAG (as a first-pass retriever, ideally paired with a reranker)
Not recommended as-is for tasks requiring fine-grained correctness judgments (e.g. automatic translation quality scoring).
License
MIT License
Model fine-tuned: 2026-07-29