modello-italia-9b-bf16/README.md

---
license: mit
language:
- it
---

# Model Card for Modello Italia 9B
This is our UNOFFICIAL conversion of the OFFICIAL model checkpoint of *"Modello Italia 9B"*, the first Large Language Model (LLM) developed by [iGenius](https://it.igenius.ai/) in collaboration [CINECA](https://www.cineca.it/).
* More information about Modello Italia: [click here](https://it.igenius.ai/language-models).

## 🚨 Disclaimers
* This is an UNOFFICIAL conversion of the OFFICIAL model checkpoint released by iGenius.
* The original model was developed using LitGPT, therefore, the weights need to be converted before they can be used with Hugging Face transformers.
* **Note:** By using this model, you accept the iGenius' [**terms and conditions**](https://secure.igenius.ai/legal/italia_terms_and_conditions.pdf).

## 🚨 Biases and Risks
From the terms and conditions of iGenius for Modello Italia:
> Modello Italia è concepito per essere utilizzato da tutti e per adattarsi a una vasta gamma di casi
  d'uso. È stato progettato con l'obiettivo di essere accessibile a persone provenienti da
  background, esperienze e prospettive diverse. Modello Italia si rivolge agli utenti e alle loro
  esigenze senza inserire giudizi superflui o normative, riconoscendo al contempo che anche
  contenuti potenzialmente problematici in determinati contesti possono avere scopi validi in altri.
  Il rispetto per la dignità e l'autonomia di tutti gli utenti, specialmente in termini di libertà di
  pensiero ed espressione, è un pilastro fondamentale del suo design. Tuttavia, essendo una nuova
  tecnologia, Modello Italia comporta rischi legati al suo utilizzo. I test condotti finora sono stati
  eseguiti in italiano e non hanno potuto coprire tutte le possibili situazioni. Pertanto, come per
  tutti gli LLM, non è possibile prevedere in anticipo gli output di Modello Italia e il modello
  potrebbe in alcuni casi generare risposte imprecise, tendenziose o altre risposte discutibili. Prima
  di utilizzare Modello Italia in qualsiasi contesto, gli sviluppatori sono fortemente incoraggiati a
  eseguire test di sicurezza e adattamento specifici per le loro applicazioni.

We are aware of the biases and potential problematic/toxic content that current pretrained large language models exhibit: more specifically, as probabilistic models of (Italian and English) languages, they reflect and amplify the biases of their training data.

For more information about this issue, please refer to our survey paper:
* [Biases in Large Language Models: Origins, Inventory, and Discussion](https://dl.acm.org/doi/full/10.1145/3597307)

## Training dataset
The following information is based on the information we could gather, that is, it is NOT official.
Please take it with a pinch of salt as we continue to study Modello Italia.
* **The training data of Modello Italia is unknown;**
* Modello Italia is probably trained on around 1T tokens of Italian text;
* We know that the training data is mostly Italian text and source code;
* We know that the training data includes text from Editoria Nazionale.

## Tokenizer
The following information is based on the information we could gather, that is, it is NOT official.
Please take it with a pinch of salt as we continue to study Modello Italia.
* The tokenizer is **vanilla SentencePiece**, probably trained from scratch on Italian data;
* Vocabulary size is 50000.

## Model architecture
The following information is based on the information we could gather, that is, it is NOT official.
Please take it with a pinch of salt as we continue to study Modello Italia.
* The model architecture is **based on GPT-NeoX**.

## Results
Modello Italia 9B has not been evaluated on standard benchmarks yet.
We will update this model card with the results soon.
* **Want to contribute to the evaluation?** Submit a pull request!

## How to use Modello Italia with Hugging Face transformers

```python
import torch
import transformers as tr

device = "cuda" if torch.cuda.is_available() else "cpu"

tokenizer = tr.AutoTokenizer.from_pretrained("sapienzanlp/modello-italia-9b-bf16")
model = tr.AutoModelForCausalLM.from_pretrained(
  "sapienzanlp/modello-italia-9b-bf16",
  device_map=device,
  torch_dtype=torch.bfloat16
)

MY_SYSTEM_PROMPT_SHORT = (
  "Tu sei Modello Italia, un modello di linguaggio naturale addestrato da iGenius."
)
prompt = "Ciao, chi sei?"
messages = [
  {"role": "system", "content": MY_SYSTEM_PROMPT_SHORT},
  {"role": "user", "content": prompt},
]
tokenized_chat = tokenizer.apply_chat_template(
  messages, tokenize=True, add_generation_prompt=True, return_tensors="pt"
).to(device)

out = model.generate(
  tokenized_chat,
  max_new_tokens=200,
  do_sample=False
)
```