lopozz/splade-bert-base-italian-xxl-uncased
sparse encoder for Italian search
1 min readupdated
This is an asymmetric inference-free SPLADE sparse encoder built on dbmdz/bert-base-italian-xxl-uncased and trained with sentence-transformers. It maps sentences and paragraphs into a high-dimensional sparse vector space (31,102 dimensions) for efficient semantic and sparse retrieval. Trained on a corpus of over 500,000 question-answer pairs, the model is tailored for Italian text.
Key features
- Built on
dbmdz/bert-base-italian-xxl-uncased - Supports a maximum sequence length of up to 512 tokens
- Generates 31,102-dimensional sparse vector representations
- Uses dot product as its similarity function
What you can do
- Set up semantic search across Italian documents in search engines
- Run sparse retrieval and match questions to answers in knowledge bases
- Plug the model into text processing pipelines via
sentence-transformers
Example
from sentence_transformers import SparseEncoder
model = SparseEncoder("lopozz/splade-bert-base-italian-xxl-uncased")
sentences = [
"Cos'è un atomo di zolfo?",
"Lo zolfo è un elemento chimico non metallico...",
]
embeddings = model.encode(sentences)
print(embeddings.shape)
Read more
DetailsMetrics on the mmarco dataset: dot_ndcg@10 is 0.7107, dot_accuracy@10 is 0.8327, and dot_mrr@10 is 0.6771.
Training dataset size: 532,751 samples (query and answer pairs).
License: apache-2.0.










Comments