OpenVINO IR model with int8 quantization of LLaMAntino-3-ANITA-8B-Inst-DPO-ITA

Model definition for LocalAI:

name: anita-llama3
backend: transformers
parameters:
  model: fakezeta/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA-ov-int8
context_size: 8192
type: OVModelForCausalLM
template:
  use_tokenizer_template: true

To run the model directly with LocalAI:

local-ai run huggingface://fakezeta/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA-ov-int8/model.yaml

aaa

"Built with Meta Llama 3".

LLaMAntino-3-ANITA-8B-Inst-DPO-ITA is a model of the LLaMAntino - Large Language Models family. The model is an instruction-tuned version of Meta-Llama-3-8b-instruct (a fine-tuned LLaMA 3 model). This model version aims to be the a Multilingual Model 🏁 (EN 🇺🇸 + ITA🇮🇹) to further fine-tuning on Specific Tasks in Italian.

The 🌟ANITA project🌟 *(Advanced Natural-based interaction for the ITAlian language)* wants to provide Italian NLP researchers with an improved model for the Italian Language 🇮🇹 use cases.

Live DEMO: https://chat.llamantino.it/
It works only with Italian connection.

Model Details

Last Update: 10/05/2024

https://github.com/marcopoli/LLaMAntino-3-ANITA

Model	HF	GGUF	EXL2
swap-uniba/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA	Link	Link	Link

Specifications

Model developers:
Ph.D. Marco Polignano - University of Bari Aldo Moro, Italy
SWAP Research Group
Variations: The model release has been supervised fine-tuning (SFT) using QLoRA 4bit, on instruction-based datasets. DPO approach over the mlabonne/orpo-dpo-mix-40k dataset is used to align with human preferences for helpfulness and safety.
Input: Models input text only.
Language: Multilingual 🏁 + Italian 🇮🇹
Output: Models generate text and code only.
Model Architecture: Llama 3 architecture.
Context length: 8K, 8192.
Library Used: Unsloth

Playground

To use the model directly, there are many ways to get started, choose one of the following ways to experience it.

Prompt Template

<|start_header_id|>system<|end_header_id|>

{ SYS Prompt }<|eot_id|><|start_header_id|>user<|end_header_id|>

{ USER Prompt }<|eot_id|><|start_header_id|>assistant<|end_header_id|>

{ ASSIST Prompt }<|eot_id|>

Transformers

For direct use with transformers, you can easily get started with the following steps.

Firstly, you need to install transformers via the command below with pip.
```
pip install -U transformers trl peft accelerate bitsandbytes
```

Right now, you can start using the model directly.

import torch
from transformers import (
    AutoModelForCausalLM,
    AutoTokenizer,
)

base_model = "swap-uniba/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA"
model = AutoModelForCausalLM.from_pretrained(
    base_model,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained(base_model)

sys = "Sei un an assistente AI per la lingua Italiana di nome LLaMAntino-3 ANITA " \
    "(Advanced Natural-based interaction for the ITAlian language)." \
    " Rispondi nella lingua usata per la domanda in modo chiaro, semplice ed esaustivo."

messages = [
    {"role": "system", "content": sys},
    {"role": "user", "content": "Chi è Carlo Magno?"}
]

#Method 1
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt", add_special_tokens=False)
for k,v in inputs.items():
    inputs[k] = v.cuda()
outputs = model.generate(**inputs, max_new_tokens=512, do_sample=True, top_p=0.9, temperature=0.6)
results = tokenizer.batch_decode(outputs)[0]
print(results)

#Method 2
import transformers
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=0.6,  #temperature for more or less creative answers
    do_sample=True,
    top_p=0.9,
)

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

Additionally, you can also use a model with 4bit quantization to reduce the required resources at least. You can start with the code below.

import torch
from transformers import (
    AutoModelForCausalLM,
    AutoTokenizer,
    BitsAndBytesConfig,
)

base_model = "swap-uniba/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA"
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=False,
)
model = AutoModelForCausalLM.from_pretrained(
    base_model,
    quantization_config=bnb_config,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained(base_model)

sys = "Sei un an assistente AI per la lingua Italiana di nome LLaMAntino-3 ANITA " \
    "(Advanced Natural-based interaction for the ITAlian language)." \
    " Rispondi nella lingua usata per la domanda in modo chiaro, semplice ed esaustivo."

messages = [
    {"role": "system", "content": sys},
    {"role": "user", "content": "Chi è Carlo Magno?"}
]

#Method 1
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt", add_special_tokens=False)
for k,v in inputs.items():
    inputs[k] = v.cuda()
outputs = model.generate(**inputs, max_new_tokens=512, do_sample=True, top_p=0.9, temperature=0.6)
results = tokenizer.batch_decode(outputs)[0]
print(results)

#Method 2
import transformers
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=0.6,  #temperature for more or less creative answers
    do_sample=True,
    top_p=0.9,
)

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

Evaluation

Open LLM Leaderboard:

Evaluated with lm-evaluation-benchmark-harness for the Open Italian LLMs Leaderboard

   lm_eval --model hf --model_args pretrained=HUGGINGFACE_MODEL_ID  --tasks hellaswag_it,arc_it  --device cuda:0 --batch_size auto:2
   lm_eval --model hf --model_args pretrained=HUGGINGFACE_MODEL_ID  --tasks m_mmlu_it --num_fewshot 5  --device cuda:0 --batch_size auto:2

Metric	Value
Avg.	0.6160
Arc_IT	0.5714
Hellaswag_IT	0.7093
MMLU_IT	0.5672

Unsloth

Unsloth, a great tool that helps us easily develop products, at a lower cost than expected.

Citation instructions

@misc{polignano2024advanced,
      title={Advanced Natural-based interaction for the ITAlian language: LLaMAntino-3-ANITA}, 
      author={Marco Polignano and Pierpaolo Basile and Giovanni Semeraro},
      year={2024},
      eprint={2405.07101},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

@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}
}

@article{llama3modelcard,
  title={Llama 3 Model Card},
  author={AI@Meta},
  year={2024},
  url = {https://github.com/meta-llama/llama3/blob/main/MODEL_CARD.md}
}

Acknowledgments

We acknowledge the support of the PNRR project FAIR - Future AI Research (PE00000013), Spoke 6 - Symbiotic AI (CUP H97G22000210007) under the NRRP MUR program funded by the NextGenerationEU. Models are built on the Leonardo supercomputer with the support of CINECA-Italian Super Computing Resource Allocation, class C project IscrC_Pro_MRS (HP10CQO70G).

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	75.12
AI2 Reasoning Challenge (25-Shot)	74.57
HellaSwag (10-Shot)	92.75
MMLU (5-Shot)	66.85
TruthfulQA (0-shot)	75.93
Winogrande (5-shot)	82.00
GSM8k (5-shot)	58.61

fakezeta
/

LLaMAntino-3-ANITA-8B-Inst-DPO-ITA-ov-int8