README.md

---
license: llama2
language:
- it
tags:
- text-generation-inference
---
<img src="https://i.ibb.co/6mHSRm3/llamantino53.jpg" alt="llamantino53" border="0" width="200px">

# LLaMAntino-2-70b-hf-UltraChat-ITA 🇮🇹 🌟
*Last Update: 02/02/2024*<br>
<hr>

## Model description

<!-- Provide a quick summary of what the model is/does. -->

**LLaMAntino-2-70b-hf-UltraChat-ITA** is a *Large Language Model (LLM)* that is an instruction-tuned version of **LLaMAntino-2-70b** (an italian-adapted **LLaMA 2 - 70B**). 
This model aims to provide Italian NLP researchers with an improved model for italian dialogue use cases.

The model was trained using *QLora* and using as training data [UltraChat](https://github.com/thunlp/ultrachat) translated to the italian language using [Argos Translate](https://pypi.org/project/argostranslate/1.4.0/). 
If you are interested in more details regarding the training procedure, you can find the code we used at the following link:
- **Repository:** https://github.com/swapUniba/LLaMAntino

**NOTICE**: the code has not been released yet, we apologize for the delay, it will be available asap!

- **Developed by:** Pierpaolo Basile, Elio Musacchio, Marco Polignano, Lucia Siciliani, Giuseppe Fiameni, Giovanni Semeraro
- **Funded by:** PNRR project FAIR - Future AI Research
- **Compute infrastructure:** [Leonardo](https://www.hpc.cineca.it/systems/hardware/leonardo/) supercomputer
- **Model type:** LLaMA-2
- **Language(s) (NLP):** Italian
- **License:** Llama 2 Community License 
- **Finetuned from model:** [swap-uniba/meta-llama/Llama-2-70b-hf](https://huggingface.co/meta-llama/Llama-2-70b-hf)

## Prompt Format

This prompt format based on the [LLaMA 2 prompt template](https://gpus.llm-utils.org/llama-2-prompt-template/) adapted to the italian language was used:

```python
" [INST] <<SYS>>\n" \
"Sei un assistente disponibile, rispettoso e onesto di nome Llamantino. " \
"Rispondi sempre nel modo più utile possibile, pur essendo sicuro. " \
"Le risposte non devono includere contenuti dannosi, non etici, razzisti, sessisti, tossici, pericolosi o illegali. " \
"Assicurati che le tue risposte siano socialmente imparziali e positive. " \
"Se una domanda non ha senso o non è coerente con i fatti, spiegane il motivo invece di rispondere in modo non corretto. " \
"Se non conosci la risposta a una domanda, non condividere informazioni false.\n" \
"<</SYS>>\n\n" \
f"{user_msg_1} [/INST] {model_answer_1} </s> <s> [INST] {user_msg_2} [/INST] {model_answer_2} </s> ... <s> [INST] {user_msg_N} [/INST] {model_answer_N} </s>"
```

We recommend using the same prompt in inference to obtain the best results!

## How to Get Started with the Model

Below you can find an example of model usage:

```python
from transformers import AutoTokenizer, AutoModelForCausalLM
import transformers
import torch
import os
os.environ["CUDA_VISIBLE_DEVICES"] = "0,1,2,3"

model = "swap-uniba/LLaMAntino-2-70b-hf-UltraChat-ITA"

tokenizer = AutoTokenizer.from_pretrained(model)
tokenizer.add_special_tokens({"pad_token":"<unk>"})
tokenizer.chat_template =   "{% set ns = namespace(i=0) %}" \
                            "{% for message in messages %}" \
                                "{% if message['role'] == 'user' and ns.i == 0 %}" \
                                       "{{ bos_token +' [INST] <<SYS>>\n' }}" \
                                       "{{ 'Sei un assistente disponibile, rispettoso e onesto di nome Llamantino. ' }}" \
                                       "{{ 'Rispondi sempre nel modo più utile possibile, pur essendo sicuro. ' }}" \
                                       "{{ 'Le risposte non devono includere contenuti dannosi, non etici, razzisti, sessisti, tossici, pericolosi o illegali. ' }}" \
                                       "{{ 'Assicurati che le tue risposte siano socialmente imparziali e positive. ' }}" \
                                       "{{ 'Se una domanda non ha senso o non è coerente con i fatti, spiegane il motivo invece di rispondere in modo non corretto. ' }}" \
                                       "{{ 'Se non conosci la risposta a una domanda, non condividere informazioni false.\n' }}" \
                                       "{{ '<</SYS>>\n\n' }}" \
                                       "{{ message['content'] + ' [/INST]' }}" \
                                "{% elif message['role'] == 'user' and ns.i != 0 %} " \
                                    "{{ bos_token + ' [INST] ' + message['content'] + ' [/INST]' }}" \
                                "{% elif message['role'] == 'assistant' %}" \
                                    "{{ ' '  + message['content'] + ' ' + eos_token + ' ' }}" \
                                "{% endif %}" \
                                "{% set ns.i = ns.i+1 %}" \
                            "{% endfor %}"

model = AutoModelForCausalLM.from_pretrained(
    model,
    torch_dtype=torch.float16,
    device_map='balanced',
    use_flash_attention_2=True
)

pipe = transformers.pipeline(model=model,
    device_map="balanced",
    tokenizer=tokenizer,
    return_full_text=False,  # langchain expects the full text
    task='text-generation',
    max_new_tokens=512,  # max number of tokens to generate in the output
    temperature=0.7 #temperature
)
messages = [{"role": "user", "content": "Cosa sono i word embeddings?"}]
text = tokenizer.apply_chat_template(messages, tokenize=False)

sequences = pipe(text)
for seq in sequences:
    print(f"{seq['generated_text']}")

```

If you are facing issues when loading the model, you can try to load it **Quantized**:

```python
model = AutoModelForCausalLM.from_pretrained(model_id, load_in_8bit=True)
```

*Note*:
1) The model loading strategy above requires the [*bitsandbytes*](https://pypi.org/project/bitsandbytes/) and [*accelerate*](https://pypi.org/project/accelerate/) libraries
2) The Tokenizer, by default, adds at the beginning of the prompt the '\<BOS\>' token. If that is not the case, add as a starting token the *\<s\>* string.
   
## Evaluation

For a detailed comparison of model performance, check out the [Leaderboard for Italian Language Models](https://huggingface.co/spaces/FinancialSupport/open_ita_llm_leaderboard).

Here's a breakdown of the performance metrics:
| Metric                      | hellaswag_it acc_norm | arc_it acc_norm | m_mmlu_it 5-shot acc | Average |
|:----------------------------|:----------------------|:----------------|:---------------------|:--------|
| **Accuracy Normalized**     | 0.6566              | 0.5004        | 0.6084                 | 0.588  |



## Citation

<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->

If you use this model in your research, please cite the following:

```bibtex
@misc{basile2023llamantino,
      title={LLaMAntino: LLaMA 2 Models for Effective Text Generation in Italian Language}, 
      author={Pierpaolo Basile and Elio Musacchio and Marco Polignano and Lucia Siciliani and Giuseppe Fiameni and Giovanni Semeraro},
      year={2023},
      eprint={2312.09993},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}
```

*Notice:* Llama 2 is licensed under the LLAMA 2 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved. [*License*](https://ai.meta.com/llama/license/)