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 Details

Training Data

Dataset used: Opus100
Subset: "en-fr"

Evaluation

global_step=5000
training_loss=1.295289501953125

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

Compute Instance

Google Colab - T4 GPU (Free)

Framework versions

PEFT 0.7.1