An extensible, convenient, and efficient toolbox for finetuning large machine learning models, designed to be user-friendly, speedy and reliable, and accessible to the entire community.
Large Model for All. See our vision.
- [2023-07-23] 🚀 LMFlow multimodal chatbot is now available! Support multimodal inputs of images and texts. Online Demo is also provided (We hold the service on a single GPU, hence one may experience "queuing" or "application busy" sometimes when multiple users are accessing at the same time, please wait and attempt again later when such event happens) 🚀
- [2023-06-22] LMFlow paper is out! Check out our implementation details at https://arxiv.org/abs/2306.12420
- [2023-06-16] Our finetuned Robin-33B-V2 scored an impressive 64.1 on the Huggingface LLM leaderboard in our offline evaluation, outperforming major open-source LLMs! All checkpoints (7B, 13B, 33B, and 65B) are released! Checkout the performance here.
- [2023-06-07] LMFlow is now officially available on PyPI! Install it with
pip install lmflow-finetune
! - [2023-05-30] Release Robin-13B-v2 and Robin-33B-v2!
- [2023-05-15] Release LMFlow-data, the training dataset of Robin-7B-v2. A new test data is also released.
- [2023-05-09] Release Robin-7B-v2, achieving competitive performance on chitchat, commonsense reasoning and instruction-following tasks. Refer to our comprehensive study.
- [2023-05-08] Release LMFlow Benchmark, an automatic evaluation framework for open-source chat-style LLMs. Benchmark results on 31 popular models are reported. Participate in LMFlow Benchmark.
- [2023-04-21] Release Robin-7B (based on LLaMA-7B), and two models for commercial use: Parakeets-2.7B (based on GPT-NEO-2.7B) and Cokatoo-7B (based on StableLM-7B) Download here
- [2023-04-10] We propose a new alignment algorithm: Reward rAnked FineTuning (RAFT), which is more efficient than conventional (PPO-based) RLHF. [Paper]
- [2023-04-02] Web service is online!
- [2023-04-01] Release three instruction-tuned checkpoints and three medical checkpoints in model zoo: LLaMA-7B-tuned, LLaMA-13B-tuned, LLaMA-33B-tuned, LLaMA-7B-medical, LLaMA-13B-medical, and LLaMA-33B-medical.
- [2023-03-27] Release code and checkpoints - version 0.0.1! Our tasked-tuned model beats ChatGPT on medical domain.
We provide four kinds of demos which include
- Online Service: If you don't want to run any code and just want to try our models, we deploy our instruction-tuned LLaMA you to have a try.
- Colab Chatbot (shell): An interactive shell-based chatbot for you to easily deploy a chatbot on colab.
- Colab Chatbot (web): An interactive web-based chatbot for you to easily deploy your own chatbot on colab.
- Local Deploy: We also provide a way for you to deploy your model/chatbot locally, which means you can deploy much larger model than previous three methods if you have enough resource.
Welcome to visit our web service. We deploy LLaMA-7B-tuned model online for preview. Due to the high website traffic, sometimes the website may fail to respond. You can also deploy the chatbot following Local Deploy
.
We provide a simple shell demo of chatbot with Google Colab's T4/P100/V100 GPU. Notice that the provided gpt-neo-2.7b model is a rather weak model, which only supports English and may sometimes generate unsatisfactory responses. To improve the performance, users can use their own dataset to finetune and obtain a better model with LMFlow. One can also try other available decoder-only models provided in 🤗 Hugging Face, by
./scripts/run_chatbot.sh {another-model-name}
We provide a simple web demo of chatbot with Google Colab's T4/P100/V100 GPU. Notice that the provided gpt-neo-2.7b model is a rather weak model, which only supports English and may sometimes generate unsatisfactory responses.
If you have resources and want to deploy your own model locally. We provide you an easy way to run a flask server to launch a backend (to further provide services to other frontend) and an interactive web frontend (to let you communicate directly) by
./scripts/run_app.sh
We also provide a gradio-based UI for building chatbots. Running the following command will launch the demo for robin-7b:
pip install gradio
python ./examples/chatbot_gradio.py --deepspeed configs/ds_config_chatbot.json --model_name_or_path YOUR-LLAMA --lora_model_path ./robin-7b --prompt_structure "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.###Human: {input_text}###Assistant:" --end_string "#" --max_new_tokens 200
We also hosted it on Hugging Face Space.
We also provide a simple demo to display the effectiveness of RAFT algorithm on diffusion models.
You can refer to either Colab link or experimental/RAFT-diffusion/SD256-RAFT.ipynb
.
The support of multi-modal training of LMFlow is under development.
PubMedQA (ID) | MedQA-USMLE (OOD) | MedMCQA (ID) | Average | |
---|---|---|---|---|
Human (pass) | 60.0 | 50.0 | ||
Human (expert) | 78.0 | 87.0 | 90.0 | 85.0 |
InstructGPT 175B | 73.2 | 46.0 | 44.0 | 54.4 |
ChatGPT | 63.9 | 57.0 | 44.7 | 55.2 |
LLaMA 7B | 5.2 | 27.1 | 24.3 | 18.9 |
LLaMA 33B | 1.8 | 43.4 | 30.3 | 25.2 |
Task-tuned LLaMA 7B (Full) | 75.1 | 44.5 | 49.9 | 56.5 |
Task-tuned LLaMA 33B (LoRA) | 74.0 | 51.3 | 50.2 | 58.5 |
The LLaMA 33B (LoRA) performance is achieved with only ~16h finetuning on the training split of PubMedQA and MedMCQA with a single 8 * A100 server. For more performance, including instruction tuning results, please refer to our Documentation.
We open-sourced the trained checkpoints to everyone for further training and inference.
![]() |
![]() |
![]() |
![]() |
|
Model Base Model |
Robin-7B ⭐ LLaMA-7B |
Robin-13B LLaMA-13B |
Robin-33B LLaMA-33B |
Robin-65B LLaMA-65B |
![]() |
![]() |
![]() |
![]() |
|
Model Base Model |
Robin-7B-medical LLaMA-7B |
Robin-13B-medical LLaMA-13B |
Robin-33B-medical LLaMA-33B |
Robin-65B-medical LLaMA-65B |
![]() |
![]() |
![]() |
||
Model Base Model |
Parakeets-2.7B ⭐ GPT-NEO-2.7B |
Cockatoo-3B StableLM-3B |
Cockatoo-7B ⭐ StableLM-7B |
Pipelines | Status |
---|---|
Task Tuning | ✅ Supported |
Instruction Tuning | ✅ Supported |
Parameter-Efficient Tuning | ✅ Supported |
Alignment Tuning | ✅ Supported |
Large Model Inference | ✅ Supported |
Seamlessly supported all the decoder models in 🤗 Hugging Face. LLaMA, GPT2, GPT-Neo, Galactica, have been fully tested. We will support encoder models soon.
Our package has been fully tested on Linux OS (Ubuntu 20.04). Other OS platforms (MacOS, Windows) are not fully tested. You may encounter some unexpected errors. You may try it first on a Linux machine or use Google Colab to experience it.
git clone https://github.com/OptimalScale/LMFlow.git
cd LMFlow
conda create -n lmflow python=3.9 -y
conda activate lmflow
conda install mpi4py
pip install -e .
Please refer to our doc.
You can run scripts/run_finetune.sh
to finetune a GPT-2 base model
./scripts/run_finetune.sh
If you would like to provide arguments for deepspeed to reflect your machine settings, you may pass the corresponding deepspeed arguments to the script. For example,
./scripts/run_finetune.sh "--num_gpus=8 --master_port 10001"
To enable LoRA finetuning, you may refer to
./scripts/run_finetune_with_lora.sh
which can be run in similar manner.
For detailed configurations, one may modify these scripts directly. These
scripts actually just call python script examples/finetune.py
, which can
be run in following manner,
deepspeed ${deepspeed_args} \
examples/finetune.py \
--deepspeed configs/ds_config_zero3.json \
--bf16 \
--run_name finetune_with_lora \
--model_name_or_path facebook/galactica-1.3b \
--num_train_epochs 0.01 \
--learning_rate 2e-5 \
--dataset_path ${dataset_path} \
--per_device_train_batch_size 1 \
--per_device_eval_batch_size 1 \
--validation_split_percentage 0 \
--logging_steps 20 \
--block_size 512 \
--do_train \
--output_dir output_models/finetune \
--overwrite_output_dir \
--ddp_timeout 72000 \
--save_steps 5000 \
--dataloader_num_workers 1
Here we set number of epochs --num_train_epochs
to 0.01
so that the
finetuning process can be finished quickly. If you wish to obtain a model with
better performance, feel free to adjust those hyperparameters. You may run
python examples/finetune.py -h
Note: In the case of a small training data set, the value of block_size
needs to be reduced, otherwise there will be no samples available in the Epoch iterator.
to view all possible finetuning arguments. The finetuned model checkpoint will
be saved in the argument specified by --output_dir
, which is
output_models/finetune
in the above example.
We follow Alpaca and Vicuna in the model tuning process and serve the model in our web service.
One can directly run evaluation with an existing Hugging Face model, e.g. to run GPT2 large, one may execute
./scripts/run_evaluation.sh
or run the corresponding python script
CUDA_VISIBLE_DEVICES=0 \
deepspeed examples/evaluate.py \
--answer_type medmcqa \
--model_name_or_path gpt2-large \
--dataset_path data/MedQA-USMLE/validation \
--deepspeed examples/ds_config.json
To load the finetuned model, specify --model_name_or_path
with the saved
model checkpoint directory path.
For LoRA finetuned models, one may refer to
./scripts/run_evaluation_with_lora.sh
Those scripts invoke the examples examples/*.py
built based on our APIs. For
more API-related examples, one may refer to the methods in the unittest
tests
.
LMFlow Benchmark is an automatic evaluation framework for open-source large language models. We use negative log likelihood (NLL) as the metric to evaluate different aspects of a language model: chitchat, commonsense reasoning, and instruction following abilities.
You can directly run the LMFlow benchmark evaluation to obtain the results to participate in the LLM comparision. For example, to run GPT2 XL, one may execute
./scripts/run_benchmark.sh --model_name_or_path gpt2-xl
--model_name_or_path
is required, you may fill in huggingface model name or local model path here.
To check the evaluation results, you may check benchmark.log
in ./output_dir/gpt2-xl_lmflow_chat_nll_eval
,
./output_dir/gpt2-xl_all_nll_eval
and ./output_dir/gpt2-xl_commonsense_qa_eval
.
Please refer to our doc.
You can config the deepspeed under configs. Details can be referred at DeepSpeed Configuration
Thanks to the great efforts of llama.cpp. It is possible for everyone to run their LLaMA models on CPU by 4-bit quantization. We provide a script to convert LLaMA LoRA weights to .pt
files. You only need to use convert-pth-to-ggml.py
in llama.cpp to perform quantization.
You can run following script to download our medical model checkpoints :
cd output_models
bash download.sh medical_ckpt
cd -
You can also directly download our model via google drive link : medical_ckpt.tar.gz
Similarly, you can run following script to download our instruction model checkpoints :
cd output_models
bash download.sh instruction_ckpt
cd -
You can also directly download our model via google drive link : instruction_ckpt.tar.gz
After downloading the model checkpoints, you can merge the lora model into the base model via
python examples/merge_lora.py \
--model_name_or_path {huggingface-model-name-or-path-to-base-model} \
--lora_model_path {path-to-lora-model} \
--output_model_path {path-to-merged-model}
Or you can replace the --lora_model_path
with output_models/instruction_ckpt/llama7b-lora
(example for llama-7b for instruction) and replace --model_name_or_path
with your converted llama model inside LMFlow/scripts/run_evaluation_with_lora.sh
and run this shell script to reproduce the result.
For full model deltas, such as robin-7b-v2-delta, you may use the delta merge script to obtain the full model,
python utils/apply_delta.py \
--base-model-path {huggingface-model-name-or-path-to-base-model} \
--delta-path {path-to-delta-model} \
--target-model-path {path-to-merged-model}
Then you can check the model performance at our Doc.
Please refer to our Documentation for more API reference and experimental results.
Hello there! We are excited to announce the upcoming release of our code repository that includes a complete LLM training process, enabling users to quickly build their own language models and train them effectively.
Our code repository is not just a simple model; it includes the complete training workflow, model optimization, and testing tools. You can use it to build various types of language models, including conversation models, question-answering models, and text generation models, among others.
Moreover, we aim to create an open and democratic LLM sharing platform where people can share their checkpoints and experiences to collectively improve the skills of the community. We welcome anyone who is interested in LLM to participate and join us in building an open and friendly community!
Whether you are a beginner or an expert, we believe that you can benefit from this platform. Let's work together to build a vibrant and innovative LLM community!
LMFlow draws inspiration from various studies, including but not limited to:
- Alpaca: https://github.com/tatsu-lab/stanford_alpaca
- Vicuna: https://github.com/lm-sys/FastChat
If you need any help, please submit a Github issue.
The code included in this project is licensed under the Apache 2.0 license. If you wish to use the codes and models included in this project for commercial purposes, please sign this document to obtain authorization.
If you find this repository useful, please consider giving ⭐ and citing our paper:
@misc{lmflow,
author = {Shizhe Diao and Rui Pan and Hanze Dong and KaShun Shum and Jipeng Zhang and Wei Xiong and Tong Zhang},
title = {LMFlow: An Extensible Toolkit for Finetuning and Inference of Large Foundation Models},
year = {2023},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {\url{https://optimalscale.github.io/LMFlow/}},
}