This repository contains code for extending the Stanford Alpaca synthetic instruction tuning to existing instruction-tuned models such as Flan-T5. The pretrained models and demos are available on HuggingFace 🤗 : Base (220M), Large (770M), XL (3B), XXL (11B, Coming soon)
Alpaca represents an exciting new direction to approximate the performance of large language models (LLMs) like ChatGPT cheaply and easily. Concretely, they leverage an LLM such as GPT-3 to generate instructions as synthetic training data. The synthetic data which covers more than 50k tasks can then be used to finetune a smaller model. However, the original implementation is less accessible due to licensing constraints of the underlying LLaMA model. Furthermore, users have noted potential noise in the synthetic dataset. Hence, it may be better to explore a fully accessible model that is already trained on high-quality (but less diverse) instructions such as Flan-T5.
from transformers import pipeline
prompt = "Write an email about an alpaca that likes flan"
model = pipeline(model="declare-lab/flan-alpaca-xl")
model(prompt, max_length=128, do_sample=True)
# Dear AlpacaFriend,
# My name is Alpaca and I'm 10 years old.
# I'm excited to announce that I'm a big fan of flan!
# We like to eat it as a snack and I believe that it can help with our overall growth.
# I'd love to hear your feedback on this idea.
# Have a great day!
# Best, AL Paca
Install dependencies and download data. We used the cleaned data from Alpaca-LoRA for training.
conda create -n paca python=3.8 -y
conda activate paca
pip install -r requirements.txt
mkdir -p data
wget https://raw.githubusercontent.com/tatsu-lab/stanford_alpaca/main/alpaca_data.json -O data/alpaca.json
wget https://raw.githubusercontent.com/tloen/alpaca-lora/main/alpaca_data_cleaned.json -O data/alpaca_clean.json
Preprocess training dataset:
python data_loading.py preprocess_alpaca \
--path_in data/alpaca_clean.json \
--path_out data/train.json
The following command will finetune the Flan-T5-XL (8hrs on a single A6000 GPU).
python training.py --output_dir outputs/model/xl \
--use_compile \
--train_epochs 3 \
--data_path data/train.json \
--model_name_or_path "google/flan-t5-xl" \
--train_batch_size 1 \
--gradient_accumulation_steps 64
If the model does not fit into memory, and you have multiple GPUs, you can
try fully-sharded data parallel by replacing --use_compile
with --use_fsdp.
python inference.py test_model \
--path "outputs/model/xl/epoch=2-step=2436.ckpt" \
--prompt "Write an email about an alpaca that likes flan"
Replace declare-lab/flan-alpaca-xl with your desired HuggingFace repo.
huggingface-cli login
python inference.py export_to_hub \
--path "outputs/model/xl/epoch=2-step=2436.ckpt" \
--repo declare-lab/flan-alpaca-xl