---
language:
  - it
library_name: transformers
tags:
  - bert
  - encoder
  - italian
  - masked-language-modeling
  - long-context
  - alibi
  - matformer
---

# AlBERTmini

**AlBERTmini** is a 95M-parameter Italian encoder model from the
**AlBERTurin** family.

It was trained from scratch on approximately 7B Italian tokens using
masked language modeling.

## Model Description

AlBERTmini is the smallest model in the AlBERTurin family of
encoder-only Transformer models for Italian.

The model incorporates several architectural improvements over the
original BERT architecture, including Pre-RMSNorm, SwiGLU activations,
ALiBi positional biases, and a mask-only pre-training objective.

AlBERTmini uses:

- 6 Transformer layers
- hidden size of 768
- 12 attention heads
- SwiGLU activations
- Pre-RMSNorm
- ALiBi positional biases
- 1,024-token training sequence length
- 20% mask-only MLM
- Muon optimizer

The model uses **gettone**, a 32,768-token BPE tokenizer optimized
for Italian and shared across the AlBERTurin model family.

The model was trained using
[Matformer](https://github.com/mrinaldi97/matformer).

## AlBERTurin Model Family

| Model | Parameters | Training Tokens |
| --- | ---: | ---: |
| **AlBERTmini** | 95M | 7B |
| [AlBERTina](https://huggingface.co/AlBERTurin/AlBERTina) | 140M | 14B |
| [AlBERTone101](https://huggingface.co/AlBERTurin/AlBERTone101) | 450M | ~101B |

## Installation

```bash
python -m pip install \
  git+https://github.com/mrinaldi97/matformer.git@alberturin-v1
```

## Usage

```python
from transformers import AutoTokenizer, AutoModelForMaskedLM

model_id = "AlBERTurin/AlBERTmini"

tokenizer = AutoTokenizer.from_pretrained(model_id)

model = AutoModelForMaskedLM.from_pretrained(
    model_id,
    trust_remote_code=True,
)
```

## Citation

If you use AlBERTmini in your research, please cite:

> Matteo Rinaldi, Marco Madeddu, Calogero Jerik Scozzaro,
> Matteo Delsanto, Daniele Paolo Radicioni, and Viviana Patti.  
> **AlBERTurin: A Fully Open Family of Italian Encoder Models with
> Modern Architectures.**  
> CLiC-it 2026.