https://huggingface.co/swap-uniba

Model Card for LLaMAntino-2-ITA

Last Update: 08/01/2024
Example of Use: Colab Notebook

---

Model description

LLaMAntino-2-ITA is a Large Language Model (LLM) family that is an italian-adapted LLaMA 2 Models. This model aims to provide Italian NLP researchers with a base model for italian dialogue use cases.

The basic models were language-adapted using QLora using as training data clean_mc4_it medium. The chat models were furter trained using QLora using as training data UltraChat translated to the italian language using Argos Translate. The **fine-tuned models ** were further trained using QLora using as training data dolly-15k-it formatted in an instruction-following style or EVALITA 2023 data.

If you are interested in more details regarding the training procedure, you can find the code we used at the following link:

• Repository: https://huggingface.co/swap-uniba

• 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 supercomputer

• Model type: LLaMA-2

• Language(s) (NLP): Italian

• License: Llama 2 Community License

Prompt Format

Basic Version

Below you can find an example of model usage:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "swap-uniba/LLaMAntino-2-7b-hf-ITA"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

prompt = "Scrivi qui un possibile prompt"

input_ids = tokenizer(prompt, return_tensors="pt").input_ids
outputs = model.generate(input_ids=input_ids)

print(tokenizer.batch_decode(outputs.detach().cpu().numpy()[:, input_ids.shape[1]:], skip_special_tokens=True)[0])

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

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

Note: The model loading strategy above requires the bitsandbytes and accelerate libraries

Chat Version

This prompt format based on the LLaMA 2 prompt template adapted to the italian language was used:

" [INST]<<SYS>>\n" \
"Sei un assistente disponibile, rispettoso e onesto. " \
"Rispondi sempre nel modo piu' 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 e' 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!

Below you can find an example of model usage:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "swap-uniba/LLaMAntino-2-chat-13b-hf-UltraChat-ITA"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

user_msg = "Ciao! Come stai?"

prompt = " [INST]<<SYS>>\n" \
         "Sei un assistente disponibile, rispettoso e onesto. " \
         "Rispondi sempre nel modo piu' 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 e' 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}[/INST]"

pipe = transformers.pipeline(
    model=model,
    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=0.8  #temperature for more or less creative answers
)

# Method 1
sequences = pipe(text)
for seq in sequences:
    print(f"{seq['generated_text']}")

# Method 2
input_ids = tokenizer(prompt, return_tensors="pt").input_ids
outputs = model.generate(input_ids=input_ids, max_length=512)
print(tokenizer.batch_decode(outputs.detach().cpu().numpy()[:, input_ids.shape[1]:], skip_special_tokens=True)[0])

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

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

Note:

1. The model loading strategy above requires the bitsandbytes and 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.

Fine-Tuned Version

This prompt format based on the Alpaca model was used for fine-tuning:

"Di seguito è riportata un'istruzione che descrive un'attività, abbinata ad un input che fornisce ulteriore informazione. " \
"Scrivi una risposta che soddisfi adeguatamente la richiesta.\n\n" \
f"### Istruzione:\n{instruction}\n\n### Input:\n{input}\n\n### Risposta:\n{response}"

If no input was present in the instruction, the following prompt was used:

"Di seguito è riportata un'istruzione che descrive un'attività. " \
"Scrivi una risposta che soddisfi adeguatamente la richiesta.\n\n" \
f"### Istruzione:\n{instruction}\n\n### Risposta:\n{response}"

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

Below you can find an example of model usage:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "swap-uniba/LLaMAntino-2-7b-hf-dolly-ITA"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

instruction_text = "Estrai i nomi propri di persona dal testo che segue"
input_text = "Marco ha incontrato Matteo per strada e hanno parlato di Mirco"

prompt = "Di seguito è riportata un'istruzione che descrive un'attività, accompagnata da un input che aggiunge ulteriore informazione. " \
        f"Scrivi una risposta che completi adeguatamente la richiesta.\n\n" \
        f"### Istruzione:\n{instruction_text}\n\n" \
        f"### Input:\n{input_text}\n\n" \
        f"### Risposta:\n"

input_ids = tokenizer(prompt, return_tensors="pt").input_ids
outputs = model.generate(input_ids=input_ids)

print(tokenizer.batch_decode(outputs.detach().cpu().numpy()[:, input_ids.shape[1]:], skip_special_tokens=True)[0])

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

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

Note: The model loading strategy above requires the bitsandbytes and accelerate libraries

Evaluation

Coming soon!

Citation

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

@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