bge-m3-vistalab
Turkish embedding model fine-tuned from BAAI/bge-m3 on the VistalabBelirtecSyntheticDataset, a grounded synthetic Turkish corpus (legal + retrieval + STS + classification), using LoRA (r=32) with CachedGISTEmbedLoss and Matryoshka representation learning (dims 1024, 256).
Results (MTEB-Turkish, calibrated vs the Mizan leaderboard)
| Metric | Score |
|---|---|
| Legal (Contracts/Regulation/Caselaw avg) | 54.51 |
| MTEB Turkish (Mean TaskType) | 64.50 |
The model targets Turkish legal retrieval; the legal score reflects performance on Turkish contract, regulation, and case-law retrieval tasks.
Usage
from sentence_transformers import SentenceTransformer
model = SentenceTransformer("UlkuTuncerKucuktas/bge-m3-vistalab")
emb = model.encode(["Türkçe bir cümle", "başka bir cümle"])
Training
- Base: BAAI/bge-m3 (XLM-RoBERTa, CLS pooling, max_seq 512)
- Adaptation: LoRA r=32, alpha=64, targets [query, key, value, dense]
- Loss: CachedGISTEmbedLoss (guide bge-m3) + Matryoshka [1024, 256]; classification MNRL; STS CoSENT
- Data: grounded synthetic (legal 19.8K, retrieval 20K, STS 7.6K, classification 14.2K) + real NLI/STS ballast + MS MARCO-TR
- Precision bf16, 2 epochs, lr 2e-4