Fine Tuning LLM

Read about fine-tuning language models using the best methods out there.

• Fine Tuning LLM
  • Fine Tuning T5-small for Eng to French Translation
    • Dataset
    • Model
    • HuggingFace repo
      • Download and try the model
      • Training metrics

Fine Tuning T5-small for Eng to French Translation

Fine tuning T5 small 8bit quantized model using LoRA technique using peft and transformers library.

Dataset

Dataset used from HuggingFace for fine-tuning, opus100, specifically the fr-en subset for French and English data.

from datasets import load_dataset
dataset = load_dataset("opus100", "en-fr")
dataset

# output
# DatasetDict({
#     test: Dataset({
#         features: ['translation'],
#         num_rows: 2000
#     })
#     train: Dataset({
#         features: ['translation'],
#         num_rows: 1000000
#     })
#     validation: Dataset({
#         features: ['translation'],
#         num_rows: 2000
#     })
# })

Model

T5-small from HuggingFace for translation.

from transformers import AutoModelForSeq2SeqLM
model_id="t5-small"
model = AutoModelForSeq2SeqLM.from_pretrained(
    model_id, 
    load_in_8bit=True, 
    device_map="auto",
)

HuggingFace repo

Checkout the fine-tuned model on huggingface trained on a free T4 google colab GPU.

Download and try the model

from peft import PeftModel, PeftConfig
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("dmedhi/eng2french-t5-small")
model = AutoModelForSeq2SeqLM.from_pretrained("t5-small")
model = PeftModel.from_pretrained(model, "dmedhi/eng2french-t5-small")

context = tokenizer(["Do you want coffee?"], return_tensors='pt')
output = model.generate(**context)
result = tokenizer.decode(output[0], skip_special_tokens=True)
print(result)

# Output
# Tu veux du café?

Training metrics

# metrics
train_runtime = 1672.4371
train_samples_per_second = 23.917
train_steps_per_second = 2.99
total_flos = 685071170273280.0
train_loss = 1.295289501953125
epoch = 20.0