Open-ChatGPT: An open-source implementation of ChatGPT

Introduction

Open-ChatGPT is a open-source library that allows you to train a hyper-personalized ChatGPT-like ai model using your own data and the least amount of compute possible.

Open-ChatGPT is a general system framework for enabling an end-to-end training experience for ChatGPT-like models. It can automatically take your favorite pre-trained large language models though an OpenAI InstructGPT style three stages to produce your very own high-quality ChatGPT-style model.

We have Impleamented RLHF (Reinforcement Learning with Human Feedback) powered by transformer library and DeepsSpeed. It supports distributed training and offloading, which can fit extremly large models.

If you like the project, please show your support by leaving a star ⭐.

News

• [2023/05] 🔥 We implement Stanford Alpaca Lora.

• [2023/05] 🔥 We implement Stanford Alpaca.

• [2023/04] We released RLHF(Reinforcement Learning with Human Feedback) Pipeline .

• [2023/03] We released the code OpenChatGPT: An Open-Source libraray to train ChatBot like ChatGPT.

Table of Contents

• Open-ChatGPT: An open-source implementation of ChatGPT
  • Introduction
  • News
  • Table of Contents
  • Data Collection
    • Instruction Datasets
    • RLHF Datasets
    • Data Preprocessing
    • Data Fomatting
  • Install
  • Instruction Fintune
    • Fine-tuning Alpaca-7B
    • Using DeepSpeed
    • Fine-tuning Alpaca-7B with Lora
  • Inference
    • No Enough Memory
  • Contributing
  • License
  • Acknowledgements
  • Citation

Data Collection

Instruction Datasets

A collection of open-source instruction tuning datasets to train (text and multi-modal) chat-based LLMs (GPT-4, ChatGPT,LLaMA,Alpaca).

Referring to this (@jianzhnie), we labeled each collected dataset according to the following rules:

(Lang)Lingual-Tags:

• EN: Instruction datasets in English
• CN: Instruction datasets in Chinese
• ML: [Multi-lingual] Instruction datasets in multiple languages

(Task)Task-Tags:

• MT: [Multi-task] Datasets containing multiple tasks
• TS: [Task-specific] Datasets tailored for specific tasks

(Gen)Generation-method:

• HG: [Human Generated Dataset] Datasets created by humans
• SI: [Self-Instruct] Datasets generated using self-instruct methods
• MIX: [Mixed Dataset] Dataset contains both human and machine generated data
• COL: [Collection of Dataset] Dataset made from a collection of other datasets

Here, we only list a small part of the instruction tuning dataset list, to find more datasets, please check out the following links: jianzhnie/awesome-instruction-datasets: A collection of open-source dataset to train instruction-following LLMs (ChatGPT,LLaMA,Alpaca).

RLHF Datasets

Instruction Tuning / Reinforcement Learning from Human Feedback (RLHF) Dataset is a key component of instruction-following LLMs such as ChatGPT. Follwing is a comprehensive list of datasets used for instruction tuning in various LLMs, making it easier for researchers and developers to access and utilize these resources.

To find more datasets, please check out the following links: jianzhnie/awesome-instruction-datasets: A collection of open-source dataset to train instruction-following LLMs (ChatGPT,LLaMA,Alpaca).

Data Preprocessing

We has developed a data preprocessing code that offers a unified interface for various large language models. This code can be used to preprocess data for a variety of purposes, such as Instruct Tuning and RLHF modeling tasks. If you're interested in learning more, please check out the following links to our prompt dataset and data utilities:

• prompt_dataset.py
• data_utils.py

Data Fomatting

In our collection, all data has been formatted using the same templates. Each sample follows the following structure:

[
{"instruction": instruction string,
"input": input string, # (may be empty)
"output": output string}
]

Install

git clone https://github.com/jianzhnie/open-chatgpt.git
pip install -r requirements.txt

PEFT

• If you would like to use LORA along with other parameter-efficient methods, please install peft as an additional dependency.

DeepSpeed

• If you want to accelerate LLM training using techniques such as pipeline parallelism, gradient checkpointing, and tensor fusion. Please install DeepSpeed.

Instruction Fintune

Fine-tuning Alpaca-7B

We fine-tune our models using standard Hugging Face training code. We fine-tune LLaMA-7B and LLaMA-13B with the following hyperparameters:

You can use the following command to train Alpaca-7B with 4 x A100 (40GB).

cd examples/alpaca/
python train_alpaca.py \
    --model_name_or_path  'decapoda-research/llama-7b-hf' \
    --data_path tatsu-lab/alpaca  \
    --output_dir work_dir/ \
    --num_train_epochs 3 \
    --per_device_train_batch_size 1 \
    --per_device_eval_batch_size 1 \
    --gradient_accumulation_steps 16 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 2000 \
    --save_total_limit 5 \
    --learning_rate 2e-5 \
    --weight_decay 0. \
    --warmup_ratio 0.03 \
    --lr_scheduler_type "cosine" \
    --logging_steps 1

Using DeepSpeed

If you meet OOM error, consider this.

Naively, fine-tuning a 7B model requires about 7 x 4 x 4 = 112 GB of VRAM. Commands given above enable parameter sharding, so no redundant model copy is stored on any GPU. If you'd like to further reduce the memory footprint, here are some options:

• Turn on CPU offload for FSDP with --fsdp "full_shard auto_wrap offload". This saves VRAM at the cost of longer runtime.
• In our experience, DeepSpeed stage-3 (with offload) can at times be more memory efficient than FSDP with offload. Here's an example to use DeepSpeed stage-3 with 4 GPUs with both parameter and optimizer offload:

pip install deepspeed
cd examples/alpaca/
torchrun --nproc_per_node=8 train_alpaca.py \
    --model_name_or_path  'decapoda-research/llama-7b-hf' \
    --data_path tatsu-lab/alpaca  \
    --output_dir work_dir/  \
    --num_train_epochs 3 \
    --per_device_train_batch_size 2 \
    --per_device_eval_batch_size 2 \
    --gradient_accumulation_steps 8 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 2000 \
    --save_total_limit 5 \
    --learning_rate 2e-5 \
    --weight_decay 0. \
    --warmup_ratio 0.03 \
    --deepspeed "scripts/ds_config_zero3_auto.json"

• LoRA fine-tunes low-rank slices of the query, key, and value embedding heads. This can reduce the total memory footprint from 112GB to about 7x4=28GB.

Fine-tuning Alpaca-7B with Lora

This part reproducing the Stanford Alpaca results using low-rank adaptation (LoRA).

To fine-tune cheaply and efficiently, we use Hugging Face's PEFT as well as Tim Dettmers' bitsandbytes.

This file contains a straightforward application of PEFT to the LLaMA model, as well as some code related to prompt construction and tokenization.

python train_alpaca_lora.py \
    --model_name_or_path  decapoda-research/llama-7b-hf  \
    --data_path tatsu-lab/alpaca  \
    --output_dir work_dir_lora/ \
    --num_train_epochs 3 \
    --per_device_train_batch_size 4 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 8 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 2000 \
    --save_total_limit 5 \
    --learning_rate 2e-5 \
    --weight_decay 0. \
    --warmup_ratio 0.03 \
    --lr_scheduler_type "cosine" \
    --logging_steps 1

Inference

This file reads the foundation model from the Hugging Face model hub and the LoRA weights from tloen/alpaca-lora-7b, and runs a Gradio interface for inference on a specified input. Users should treat this as example code for the use of the model, and modify it as needed.

Example usage:

python generate_server.py \
    --model_name_or_path decapoda-research/llama-7b-hf \
    --lora_model_name_or_path  tloen/alpaca-lora-7b

No Enough Memory

If you do not have enough memory, you can enable 8-bit compression by adding --load-8bit to commands above. This can reduce memory usage by around half with slightly degraded model quality. It is compatible with the CPU, GPU, and Metal backend. Alpaca-7B with 8-bit compression can run on a single NVIDIA 3090/4080/T4/V100(16GB) GPU.

python generate_server.py \
    --model_name_or_path decapoda-research/llama-7b-hf \
    --lora_model_name_or_path  tloen/alpaca-lora-7b \
    --load_8bit

Contributing

Our purpose is to make this repo even better. If you are interested in contributing, please refer to HERE for instructions in contribution.

License

Openn-ChatGPT is released under the Apache 2.0 license.

Acknowledgements

We appreciate the work by many open-source contributors, especially:

• Alpaca-LoRA
• LoRA
• Stanford Alpaca
• Hugging Face
• LLaMa
• Vicuna

Citation

Please cite the repo if you use the data or code in this repo.

@misc{open-chatgpt,
  author = {jianzhnie},
  title = {Open-ChatGPT, a chatbot based on Llama model},
  year = {2023},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/jianzhnie/open-chatgpt}},
}