diegobit/llama-3-8b-ita-4k-orpo-v3

Model Card for Model ID

This is llama-3-8b ORPO finetuning for the italian language over a concatenation of two datasets:

• mii-community/ultrafeedback-preferences-translated-ita
• efederici/alpaca-vs-alpaca-orpo-dpo

The other two differences with diegobit/llama-3-8b-Instruct-bnb-4bit-ita-orpo are:

• the starting model, not instruct, astronomer/Llama-3-8B-Special-Tokens-Adjusted instead of unsloth/llama-3-8b-Instruct-bnb-4bit
• no loading in 4bits
• given the increased need of GPU memory, the sequence max length used for finetuning is 4096

Model Details

Model Description

• Developed by: Diego Giorgini
• Funded by: AI Technologies SRL - www.aitechnologies.it
• Language(s) (NLP): Italian
• License: llama3
• Finetuned from model: astronomer/Llama-3-8B-Special-Tokens-Adjusted

Training Details

Environment

unsloth: 2024.5
torch: 2.2

Training Data

• mii-community/ultrafeedback-preferences-translated-ita is a selection of 55k rows of the ultrafeedback dataset, translated into italian with argotranslate.
• efederici/alpaca-vs-alpaca-orpo-dpo: The Alpaca vs. Alpaca dataset is a curated blend of the Alpaca dataset and the Alpaca GPT-4 dataset, both available on HuggingFace Datasets. It uses the standard GPT dataset as the 'rejected' answer, steering the model towards the GPT-4 answer, which is considered as the 'chosen' one.

Training Procedure

Preprocessing [optional]

• No preprocessing has been performed, except for formatting with the llama3 chat_template from unsloth:

  tokenizer = get_chat_template(tokenizer, chat_template = "llama-3")

Training Hyperparameters

• Training regime: bf16

• Model loading parameters:

max_seq_length = 4096
dtype = None
load_in_4bit = False

• PEFT parameters:

r = 64  
lora_alpha = 64  
lora_dropout = 0  
bias = "none"  
random_state = 3407  
use_rslora = False  
loftq_config = None

• ORPOConfig parameters:

max_length = 4096  
max_prompt_length = max_seq_length//2  
max_completion_length = max_seq_length//2  
warmup_ratio = 0.1  
weight_decay = 0.01  
per_device_train_batch_size = 1  
gradient_accumulation_steps = 16  
learning_rate=8e-6  
beta = 0.1  
optim = "paged_adamw_8bit"  
lr_scheduler_type = "linear"  
num_train_epochs = 1

Speeds, Sizes, Times

19h on an A100-40GB

Model Card Contact

diego.giorgini@icloud.com