peftfactory 0.9.4.4

Unified Parameter-Efficient Fine-Tuning of 100+ LLMs

---

Parameter-Efficient Fine-Tuning made easy

PEFT-Factory is a fork of LLaMa-Factory ❤️, upgraded with easy to use PEFT interface, support for HuggingFace PEFT methods and datasets for benchmarking PEFT.

Installation

⚠️ Currently, the best way is to install from this repository. This will change soon.

Clone the repository

git clone [e-mail]:Wicwik/PEFT-Factory.git

Build the wheel

make build

Install wiht pip

pip install dist/llamafactory-0.9.4.dev0-py3-none-any.whl

Quickstart

Quick start of prefix tuning with LLaMa3-8B-Instruct for CB dataset. We are using the template from examples directory, but feel free to use your own config.yml.

Create some variables for envsubst

TIMESTAMP=`date +%s`
OUTPUT_DIR="saves/prefix-tuning/llama-3-8b-instruct/train_cb_${TIMESTAMP}"
DATASET="cb"
SEED=123
WANDB_PROJECT="peft-factory-train-prefix-tuning"
WANDB_NAME="prefix-tuning_llama-3-8b-instruct_train_cb"

mkdir -p "${OUTPUT_DIR}"

export OUTPUT_DIR DATASET SEED WANDB_PROJECT WANDB_NAME

Use the template

Utility envsubst replaces the occurances of env variables with their values (see the template).

envsubst < examples/peft/prefix-tuning/llama-3-8b-instruct/train.yaml > ${OUTPUT_DIR}/train.yaml

Run the factory

llamafactory-cli train ${OUTPUT_DIR}/train.yaml

---

Supported methods

PEFT method name	Support
LoRA (including variants)	✅ 🦙
Prefix Tuning	✅ 🤗
Prompt Tuning	✅ 🤗
P-Tuning	✅ 🤗
LNTuning	✅ 🤗
SVD	✅ ⚙️
BitFit	✅ ⚙️