LeonardPuettmann/LlaMaestra-3.2-1B-Instruct-v0.1-4bit

LlaMaestra - A tiny Llama model tuned for text translation

 _     _      ___  ___                _             
| |   | |     |  \/  |               | |            
| |   | | __ _| .  . | __ _  ___  ___| |_ _ __ __ _ 
| |   | |/ _` | |\/| |/ _` |/ _ \/ __| __| '__/ _` |
| |___| | (_| | |  | | (_| |  __/\__ \ |_| | | (_| |
\_____/_|\__,_\_|  |_/\__,_|\___||___/\__|_|  \__,_|

Model Card

This model was finetuned with roughly 300.000 examples of translations from English to Italian and Italian to English. The model was finetuned in a way to more directly provide a translation without much explanation.

Finetuning took about 10 hours on an A10G Nvidia GPU.

Due to its size, the model runs very well on CPUs.

Usage

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "LeonardPuettmann/LlaMaestra-3.2-1B-Instruct-v0.1-4bit"

model = AutoModelForCausalLM.from_pretrained(
    model_id, 
    device_map="auto",
    trust_remote_code=True,
)

tokenizer = AutoTokenizer.from_pretrained(model_id, add_bos_token=True, trust_remote_code=True)

row_json = [
    {"role": "system", "content": "Your job is to return translations for sentences or words from either Italian to English or English to Italian."},
    {"role": "user", "content": "Do you sell tickets for the bus?"},
]

prompt =  tokenizer.apply_chat_template(row_json, tokenize=False)
model_input = tokenizer(prompt, return_tensors="pt").to("cuda")

with torch.no_grad():
    print(tokenizer.decode(model.generate(**model_input, max_new_tokens=1024)[0]))

Data used

The source for the data were sentence pairs from tatoeba.com. The data can be downloaded from here: https://tatoeba.org/downloads

Credits

Base model: unsloth/Llama-3.2-1B-Instruct derived from meta-llama/Llama-3.2-1B-Instruct Finetuned by: Leonard Püttmann https://www.linkedin.com/in/leonard-p%C3%BCttmann-4648231a9/