Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
|
@@ -1,13 +1,59 @@
|
|
| 1 |
---
|
| 2 |
-
|
| 3 |
-
-
|
| 4 |
-
- custom-model
|
| 5 |
library_name: transformers
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 6 |
---
|
| 7 |
|
| 8 |
-
#
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 9 |
|
| 10 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 11 |
|
| 12 |
## Installation
|
| 13 |
|
|
@@ -18,11 +64,24 @@ pip install git+https://github.com/mrinaldi97/matformer.git
|
|
| 18 |
## Usage
|
| 19 |
|
| 20 |
```python
|
| 21 |
-
import
|
| 22 |
-
|
|
|
|
|
|
|
|
|
|
| 23 |
|
| 24 |
model = AutoModelForMaskedLM.from_pretrained(
|
| 25 |
-
|
| 26 |
-
trust_remote_code=True
|
| 27 |
)
|
| 28 |
```
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
+
language:
|
| 3 |
+
- it
|
|
|
|
| 4 |
library_name: transformers
|
| 5 |
+
tags:
|
| 6 |
+
- bert
|
| 7 |
+
- encoder
|
| 8 |
+
- italian
|
| 9 |
+
- masked-language-modeling
|
| 10 |
+
- long-context
|
| 11 |
+
- alibi
|
| 12 |
+
- matformer
|
| 13 |
---
|
| 14 |
|
| 15 |
+
# AlBERTmini
|
| 16 |
+
|
| 17 |
+
**AlBERTmini** is a 95M-parameter Italian encoder model from the
|
| 18 |
+
**AlBERTurin** family.
|
| 19 |
+
|
| 20 |
+
It was trained from scratch on approximately 7B Italian tokens using
|
| 21 |
+
masked language modeling.
|
| 22 |
+
|
| 23 |
+
## Model Description
|
| 24 |
+
|
| 25 |
+
AlBERTmini is the smallest model in the AlBERTurin family of
|
| 26 |
+
encoder-only Transformer models for Italian.
|
| 27 |
+
|
| 28 |
+
The model incorporates several architectural improvements over the
|
| 29 |
+
original BERT architecture, including Pre-RMSNorm, SwiGLU activations,
|
| 30 |
+
ALiBi positional biases, and a mask-only pre-training objective.
|
| 31 |
+
|
| 32 |
+
AlBERTmini uses:
|
| 33 |
+
|
| 34 |
+
- 6 Transformer layers
|
| 35 |
+
- hidden size of 768
|
| 36 |
+
- 12 attention heads
|
| 37 |
+
- SwiGLU activations
|
| 38 |
+
- Pre-RMSNorm
|
| 39 |
+
- ALiBi positional biases
|
| 40 |
+
- 1,024-token training sequence length
|
| 41 |
+
- 20% mask-only MLM
|
| 42 |
+
- Muon optimizer
|
| 43 |
|
| 44 |
+
The model uses **gettone**, a 32,768-token BPE tokenizer optimized
|
| 45 |
+
for Italian and shared across the AlBERTurin model family.
|
| 46 |
+
|
| 47 |
+
The model was trained using
|
| 48 |
+
[Matformer](https://github.com/mrinaldi97/matformer).
|
| 49 |
+
|
| 50 |
+
## AlBERTurin Model Family
|
| 51 |
+
|
| 52 |
+
| Model | Parameters | Training Tokens |
|
| 53 |
+
| --- | ---: | ---: |
|
| 54 |
+
| **AlBERTmini** | 95M | 7B |
|
| 55 |
+
| [AlBERTina](https://huggingface.co/AlBERTurin/AlBERTina) | 140M | 14B |
|
| 56 |
+
| [AlBERTone101](https://huggingface.co/AlBERTurin/AlBERTone101) | 450M | ~101B |
|
| 57 |
|
| 58 |
## Installation
|
| 59 |
|
|
|
|
| 64 |
## Usage
|
| 65 |
|
| 66 |
```python
|
| 67 |
+
from transformers import AutoTokenizer, AutoModelForMaskedLM
|
| 68 |
+
|
| 69 |
+
model_id = "AlBERTurin/AlBERTmini"
|
| 70 |
+
|
| 71 |
+
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
| 72 |
|
| 73 |
model = AutoModelForMaskedLM.from_pretrained(
|
| 74 |
+
model_id,
|
| 75 |
+
trust_remote_code=True,
|
| 76 |
)
|
| 77 |
```
|
| 78 |
+
|
| 79 |
+
## Citation
|
| 80 |
+
|
| 81 |
+
If you use AlBERTmini in your research, please cite:
|
| 82 |
+
|
| 83 |
+
> Matteo Rinaldi, Marco Madeddu, Calogero Jerik Scozzaro,
|
| 84 |
+
> Matteo Delsanto, Daniele Paolo Radicioni, and Viviana Patti.
|
| 85 |
+
> **AlBERTurin: A Fully Open Family of Italian Encoder Models with
|
| 86 |
+
> Modern Architectures.**
|
| 87 |
+
> CLiC-it 2026.
|