Vietnamese Product Granite 97M
Fine-tuned SentenceTransformer model for asymmetric Vietnamese product retrieval: short search queries are matched to longer product documents.
Base model
ibm-granite/granite-embedding-97m-multilingual-r2
Usage
from sentence_transformers import SentenceTransformer
model = SentenceTransformer("trantuan1701/product-granite-97m-vietnamese")
query_embeddings = model.encode(
["đèn led downlight at02 đổi màu"],
normalize_embeddings=True,
)
document_embeddings = model.encode(
["product document text..."],
normalize_embeddings=True,
)
Important: this Granite model was trained without query: or passage:
prefixes.
Training summary
- Loss: MultipleNegativesRankingLoss
- Max sequence length: 512
- Dataset: 15,000 query-document pairs, 1,000 product documents
- Validation split: deterministic by
doc_id
Limitations
Current validation loss is a training diagnostic only. A full retrieval benchmark should report Recall@K, MRR, and NDCG before production selection.