File size: 4,030 Bytes

---
license: cc-by-nc-sa-3.0
language:
- it
pipeline_tag: text-generation
---
--------------------------------------------------------------------------------------------------

<body>
<span class="vertical-text" style="background-color:lightgreen;border-radius: 3px;padding: 3px;"> </span>
<br>
<span class="vertical-text" style="background-color:orange;border-radius: 3px;padding: 3px;">  </span>
<br>
<span class="vertical-text" style="background-color:lightblue;border-radius: 3px;padding: 3px;">    Model: PRIMO ⭐</span>
<br>
<span class="vertical-text" style="background-color:tomato;border-radius: 3px;padding: 3px;">    Lang: IT</span>
<br>
<span class="vertical-text" style="background-color:lightgrey;border-radius: 3px;padding: 3px;">  </span>
<br>
<span class="vertical-text" style="background-color:#CF9FFF;border-radius: 3px;padding: 3px;"> </span>
</body>

--------------------------------------------------------------------------------------------------

<h3>Model description</h3>

This model is a <b>causal</b> language model for the <b>Italian</b> language, based on a GPT-like <b>[1]</b> architecture (more specifically, the model has been obtained by modifying Meta's XGLM architecture <b>[2]</b> and exploiting its 7.5B checkpoint).

The model has ~6.6B parameters and a vocabulary of 50.335 tokens. It's an instruction-based model, trained with low-rank adaptation, and it's mainly suitable for general purpose prompt-based tasks involving natural language inputs and outputs.

<h3>Example</h3>

This is an example of intended use of the model:

![example](primo_screen.png)

<h3>Quantization</h3>

The released checkpoint is quantized in 8-bit, so that it can easily be loaded and used for training and inference on ordinary hardware like consumer GPUs, and it requires the installation of the <b>transformers</b> library version >= 4.30.1 and the <b>bitsandbytes</b> library, version >= 0.37.2

On Windows operating systems, the <b>bitsandbytes-windows</b> module also needs to be installed on top. However, it appears that the module is not yet updated with some recent features, like the possibility to save the 8-bit quantized models.
In order to include this, you can install the fork in [this repo](https://github.com/francesco-russo-githubber/bitsandbytes-windows), using:

```bash
pip install git+https://github.com/francesco-russo-githubber/bitsandbytes-windows.git
```

<h3>Training</h3>

The model has been fine-tuned on a dataset of ~100K Italian prompt-response pairs, using low-rank adaptation, which took about 3 days on a NVIDIA GeForce RTX 3060 GPU. The training has been carried out with a batch size of 10 and a constant learning rate of 1e-5, exploiting gradient accumulation and gradient checkpointing in order to manage the considerable memory consumption.

<h3>Quick usage</h3>

In order to use the model for inference, it's strongly recommended to build a Gradio interface using the [code provided in this repo](https://huggingface.co/osiria/primo/blob/main/app.py), or [this repo](https://huggingface.co/osiria/primo/blob/main/app_it.py) if you want an Italian interface.

<h3>Limitations</h3>

The model might behave erratically when presented with prompts which are too far away from its pre-training or fine-tuning and, because of the probabilistic nature of its generation mechanism, it might occasionally produce biased or offensive content with respect to gender, race, ideologies, and political or religious beliefs.
These limitations imply that the model and its outputs should be used with caution, and should not be involved in situations that require the generated text to be fair or true.

<h3>References</h3>

[1] https://arxiv.org/abs/2005.14165

[2] https://arxiv.org/abs/2112.10668

<h3>License</h3>

The model is released under <b>CC-BY-NC-SA-3.0</b> license, which means it is NOT available for commercial use, but only for research purposes. This is due to the experimental nature of the model and the license of the training data.