Saba ๐Ÿ“œ๐Ÿ–‹๐Ÿ‡ฎ๐Ÿ‡น

About the model

Saba is a BERT model for Italian poetry. It was obtained via continued pretraining of dbmdz/bert-base-italian-xxl-cased on ~40k Italian song lyrics from Wikisource and Biblioteca Italiana. The objective was Masked Language Modeling (MLM).

The training code is available on GitHub.

Evaluation

The base model and the adapted model were tested on a held-out set of ~1k poems with the following results:

Model MLM Loss Perplexity
Base 3.39 29.56
Saba 1.94 6.94

Evaluation of the learned representations will be made available in the future, once a suitable dataset has been created / identified.

Why Saba?

Following the tradition of giving Italian names to BERT models for the Italian language (see AlBERTo, GilBERTo, UmBERTo), we dedicate this model to the Italian poet and novelist Umberto Saba (9 March 1883 โ€“ 25 August 1957).

Downloads last month
19
Safetensors
Model size
111M params
Tensor type
F32
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for mattiaferrarini/saba

Finetuned
(8)
this model

Dataset used to train mattiaferrarini/saba