t5-ladino-espanol

t5-ladino-espanol

This model translates from modern Spanish into Judeo-Spanish (Ladino), a historical language of the Sephardic Jewish community.

This model is a fine-tuned version of google-t5/t5-small trained on the collectivat/una-fraza-al-diya dataset, a multilingual corpus designed to support the documentation and preservation of Judeo-Spanish (Ladino), an endangered language spoken historically by Sephardic Jewish communities.

It achieves the following results on the evaluation set:

• Loss: 3.3840
• BLEU: 0.0
• Generated Length: 5.0 tokens

Model description

This model is based on the T5 architecture and was fine-tuned for a sequence-to-sequence translation task. The goal is to generate translations from Spanish into Ladino, using a small parallel corpus of aligned phrases.

Intended uses & limitations

The model is intended for:

• Educational or cultural projects related to the Judeo-Spanish language.
• Language preservation and revitalization efforts.
• Demonstration of machine translation capabilities for low-resource and endangered languages.

Limitations:

• The model was trained on a very small dataset (only 307 sentence pairs).
• It may produce short or incomplete translations.
• Orthographic variation is expected, as Ladino does not have a standardized modern spelling.

Training and evaluation data

The training data comes from the dataset collectivat/una-fraza-al-diya, which contains 307 aligned phrases in Ladino, Spanish, Turkish, and English. The dataset was developed by the Sephardic Center of Istanbul as part of a cultural preservation initiative. Only the Spanish-Ladino pairs were used for training this model.

The dataset was split into:

• Training set: 245 examples (80%)
• Validation set: 31 examples (10%)
• Test set: 31 examples (10%)

Training procedure

The model was fine-tuned using the Seq2SeqTrainer class from Hugging Face's transformers library.

Training hyperparameters

The following hyperparameters were used:

• learning_rate: 5.6e-05
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• optimizer: AdamW (betas=(0.9, 0.999), epsilon=1e-08)
• lr_scheduler_type: linear
• num_epochs: 2

Training results

Framework versions

• Transformers: 4.49.0
• PyTorch: 2.6.0+cu124
• Datasets: 3.4.1
• Tokenizers: 0.21.1

Training results

Framework versions

• Transformers 4.49.0
• Pytorch 2.6.0+cu124
• Datasets 3.4.1
• Tokenizers 0.21.1