---
language:
- it
datasets:
- oscar
tags:
 - seq2seq
 - lm-head
license: apache-2.0
inference: false
---

# Italian T5-base 🇮🇹

Created by [Gabriele Sarti](https://gsarti.com/) during the [Hugging Face community week](https://discuss.huggingface.co/t/open-to-the-community-community-week-using-jax-flax-for-nlp-cv/7104), organized by [HuggingFace](https://huggingface.co/) and TPU usage sponsored by Google, for the project [PreTrain T5 for Italian](https://discuss.huggingface.co/t/pretrain-t5-for-italian/7425/4).

This is notably the first sequence-to-sequence model pre-trained on the Italian language available on the 🤗 Hub. For people interested in studying the pre-training dynamics of this model, the repository [`t5-base-it-training`](https://huggingface.co/gsarti/t5-base-it-training/tree/main) contains Flax checkpoints for the whole pre-training process (saved each 2000 steps, 129 checkpoints, ~250GB).

**Important:** The inference widget is deactivated because the model needs a task-specific seq2seq training on a downstream task to be actually useful. The script [`run_t5_mlm_flax.py`](https://github.com/huggingface/transformers/tree/master/examples/flax/language-modeling) provides an example of fine-tuning the model on a downstream summarization task.

## Dataset

This model was trained on the Italian de-duplicated portion of the [OSCAR corpus](https://oscar-corpus.com/) (11B words, ~69GB) using the 🤗 Datasets library. The corpus was used as-is without any further preprocessing.

## Training

The model was trained for 258K steps in 4 days using JAX/Flax on a TPU3v8-VM on Google Cloud. Refer to the Tensorboard tab of the repository for an overview of the training process.

The original configuration for the model `t5-base` was adopted, with the exception of the parameter `dropout_rate` that was set at `0` instead of `0.1` during pre-training, following the implementation of [`t5-v1.1`](https://huggingface.co/google/t5-v1_1-base). The tokenizer is a `SentencePieceUnigramTokenizer` trained on the first 2M sentences of the Italian portion of the [`mC4`](https://huggingface.co/datasets/mc4) corpus.

The following parameters were used for training:

|parameter|value|
|---------|-----|
|optimizer|adafactor w/ default params|
|dataset| `oscar/unshuffled_deduplicated_it`|
|max seq. length| 512|
|per-device batch size| 16|
|tot. batch size| 128|
|learning rate| 1e-2|
|lr schedule| linear warmup + linear decay|
|warmup steps|10K|
|weight decay|1e-3|
|num. train epochs| 1 (258K steps)|
|validation split size| 15K examples|