2023-12-02 01:31:24 +08:00
2023-05-28 18:09:04 +08:00
2023-10-12 21:42:29 +08:00
[](https://github.com/hiyouga/LLaMA-Factory/stargazers)
[](LICENSE)
[](https://github.com/hiyouga/LLaMA-Factory/commits/main)
2023-07-15 17:20:39 +08:00
[](https://pypi.org/project/llmtuner/)
2023-09-16 17:33:01 +08:00
[](https://pypi.org/project/llmtuner/)
2023-10-12 21:42:29 +08:00
[](https://github.com/hiyouga/LLaMA-Factory/pulls)
2023-12-15 22:11:31 +08:00
[](https://discord.gg/rKfvV9r9FK)
2023-11-16 15:58:37 +08:00
[](https://huggingface.co/spaces/hiyouga/LLaMA-Board)
2023-11-21 13:15:46 +08:00
[](https://modelscope.cn/studios/hiyouga/LLaMA-Board)
2023-05-29 21:53:02 +08:00
2023-06-02 21:47:10 +08:00
👋 Join our [WeChat ](assets/wechat.jpg ).
2023-07-21 16:57:58 +08:00
\[ English | [中文 ](README_zh.md ) \]
2023-10-16 00:23:37 +08:00
## LLaMA Board: A One-stop Web UI for Getting Started with LLaMA Factory
2023-10-15 20:28:14 +08:00
2023-11-21 13:15:46 +08:00
Preview LLaMA Board at ** [🤗 Spaces ](https://huggingface.co/spaces/hiyouga/LLaMA-Board )** or ** [ModelScope ](https://modelscope.cn/studios/hiyouga/LLaMA-Board )**.
2023-11-15 23:51:26 +08:00
2023-11-18 11:09:52 +08:00
Launch LLaMA Board via `CUDA_VISIBLE_DEVICES=0 python src/train_web.py` . (multiple GPUs are not supported yet in this mode)
2023-10-16 00:23:37 +08:00
Here is an example of altering the self-cognition of an instruction-tuned language model within 10 minutes on a single GPU.
2023-10-15 20:23:22 +08:00
https://github.com/hiyouga/LLaMA-Factory/assets/16256802/6ba60acc-e2e2-4bec-b846-2d88920d5ba1
2023-11-18 11:09:52 +08:00
## Table of Contents
- [Benchmark ](#benchmark )
- [Changelog ](#changelog )
- [Supported Models ](#supported-models )
- [Supported Training Approaches ](#supported-training-approaches )
- [Provided Datasets ](#provided-datasets )
- [Requirement ](#requirement )
- [Getting Started ](#getting-started )
- [Projects using LLaMA Factory ](#projects-using-llama-factory )
- [License ](#license )
- [Citation ](#citation )
- [Acknowledgement ](#acknowledgement )
## Benchmark
2023-11-18 11:30:01 +08:00
Compared to ChatGLM's [P-Tuning ](https://github.com/THUDM/ChatGLM2-6B/tree/main/ptuning ), LLaMA-Factory's LoRA tuning offers up to **3.7 times faster** training speed with a better Rouge score on the advertising text generation task. By leveraging 4-bit quantization technique, LLaMA-Factory's QLoRA further improves the efficiency regarding the GPU memory.
2023-11-18 11:09:52 +08:00
2023-12-01 22:53:15 +08:00
< details > < summary > Definitions< / summary >
2023-11-18 11:15:56 +08:00
- **Training Speed**: the number of training samples processed per second during the training. (bs=4, cutoff_len=1024)
2023-11-18 11:30:01 +08:00
- **Rouge Score**: Rouge-2 score on the development set of the [advertising text generation ](https://aclanthology.org/D19-1321.pdf ) task. (bs=4, cutoff_len=1024)
2023-11-18 11:15:56 +08:00
- **GPU Memory**: Peak GPU memory usage in 4-bit quantized training. (bs=1, cutoff_len=1024)
2023-11-18 11:09:52 +08:00
- We adopt `pre_seq_len=128` for ChatGLM's P-Tuning and `lora_rank=32` for LLaMA-Factory's LoRA tuning.
2023-12-01 22:53:15 +08:00
< / details >
2023-05-31 16:54:06 +08:00
## Changelog
2023-12-23 02:17:41 +08:00
[23/12/23] We supported ** [unsloth ](https://github.com/unslothai/unsloth )**'s implementation to boost LoRA tuning for the LLaMA, Mistral and Yi models. Try `--use_unsloth` argument to activate unsloth patch. It achieves 1.7x speed in our benchmark, check [this page ](https://github.com/hiyouga/LLaMA-Factory/wiki/Performance-comparison ) for details.
2023-12-23 00:14:33 +08:00
2023-12-12 11:44:30 +08:00
[23/12/12] We supported fine-tuning the latest MoE model ** [Mixtral 8x7B ](https://huggingface.co/mistralai/Mixtral-8x7B-v0.1 )** in our framework. See hardware requirement [here ](#hardware-requirement ).
2023-12-12 11:39:04 +08:00
2023-12-12 19:45:59 +08:00
[23/12/01] We supported downloading pre-trained models and datasets from the ** [ModelScope Hub ](https://modelscope.cn/models )** for Chinese mainland users. See [this tutorial ](#use-modelscope-hub-optional ) for usage.
2023-12-01 16:11:30 +08:00
2023-12-01 22:53:15 +08:00
< details > < summary > Full Changelog< / summary >
2023-12-12 11:44:30 +08:00
[23/10/21] We supported ** [NEFTune ](https://arxiv.org/abs/2310.05914 )** trick for fine-tuning. Try `--neftune_noise_alpha` argument to activate NEFTune, e.g., `--neftune_noise_alpha 5` .
2023-09-28 14:39:16 +08:00
[23/09/27] We supported ** $S^2$-Attn** proposed by [LongLoRA ](https://github.com/dvlab-research/LongLoRA ) for the LLaMA models. Try `--shift_attn` argument to enable shift short attention.
2023-09-27 21:55:50 +08:00
2023-09-23 21:10:17 +08:00
[23/09/23] We integrated MMLU, C-Eval and CMMLU benchmarks in this repo. See [this example ](#evaluation ) to evaluate your models.
2023-09-10 20:43:56 +08:00
2023-12-12 22:47:06 +08:00
[23/09/10] We supported ** [FlashAttention-2 ](https://github.com/Dao-AILab/flash-attention )**. Try `--flash_attn` argument to enable FlashAttention-2 if you are using RTX4090, A100 or H100 GPUs.
2023-08-18 01:41:17 +08:00
2023-09-23 00:34:17 +08:00
[23/08/12] We supported **RoPE scaling** to extend the context length of the LLaMA models. Try `--rope_scaling linear` argument in training and `--rope_scaling dynamic` argument at inference to extrapolate the position embeddings.
2023-08-11 03:02:53 +08:00
2023-09-23 00:34:17 +08:00
[23/08/11] We supported ** [DPO training ](https://arxiv.org/abs/2305.18290 )** for instruction-tuned models. See [this example ](#dpo-training ) to train your models.
2023-07-31 23:42:32 +08:00
2023-09-23 00:34:17 +08:00
[23/07/31] We supported **dataset streaming** . Try `--streaming` and `--max_steps 10000` arguments to load your dataset in streaming mode.
2023-08-01 10:08:47 +08:00
2023-09-23 00:34:17 +08:00
[23/07/29] We released two instruction-tuned 13B models at Hugging Face. See these Hugging Face Repos ([LLaMA-2](https://huggingface.co/hiyouga/Llama-2-Chinese-13b-chat) / [Baichuan ](https://huggingface.co/hiyouga/Baichuan-13B-sft )) for details.
2023-07-18 00:18:25 +08:00
2023-09-23 00:34:17 +08:00
[23/07/18] We developed an **all-in-one Web UI** for training, evaluation and inference. Try `train_web.py` to fine-tune models in your Web browser. Thank [@KanadeSiina ](https://github.com/KanadeSiina ) and [@codemayq ](https://github.com/codemayq ) for their efforts in the development.
2023-07-09 14:57:13 +08:00
2023-09-23 00:34:17 +08:00
[23/07/09] We released ** [FastEdit ](https://github.com/hiyouga/FastEdit )** ⚡🩹, an easy-to-use package for editing the factual knowledge of large language models efficiently. Please follow [FastEdit ](https://github.com/hiyouga/FastEdit ) if you are interested.
2023-06-29 19:36:22 +08:00
2023-09-23 00:34:17 +08:00
[23/06/29] We provided a **reproducible example** of training a chat model using instruction-following datasets, see [Baichuan-7B-sft ](https://huggingface.co/hiyouga/Baichuan-7B-sft ) for details.
2023-06-23 00:17:05 +08:00
2023-09-23 00:34:17 +08:00
[23/06/22] We aligned the [demo API ](src/api_demo.py ) with the [OpenAI's ](https://platform.openai.com/docs/api-reference/chat ) format where you can insert the fine-tuned model in **arbitrary ChatGPT-based applications** .
[23/06/03] We supported quantized training and inference (aka ** [QLoRA ](https://github.com/artidoro/qlora )**). Try `--quantization_bit 4/8` argument to work with quantized models.
2023-06-04 00:08:56 +08:00
2023-12-01 22:53:15 +08:00
< / details >
2023-05-31 16:54:06 +08:00
## Supported Models
2023-08-07 15:02:02 +08:00
2023-09-06 21:43:06 +08:00
| Model | Model size | Default module | Template |
| -------------------------------------------------------- | --------------------------- | ----------------- | --------- |
2023-12-18 22:29:45 +08:00
| [Baichuan ](https://huggingface.co/baichuan-inc ) | 7B/13B | W_pack | baichuan |
| [Baichuan2 ](https://huggingface.co/baichuan-inc ) | 7B/13B | W_pack | baichuan2 |
2023-11-06 17:18:12 +08:00
| [BLOOM ](https://huggingface.co/bigscience/bloom ) | 560M/1.1B/1.7B/3B/7.1B/176B | query_key_value | - |
| [BLOOMZ ](https://huggingface.co/bigscience/bloomz ) | 560M/1.1B/1.7B/3B/7.1B/176B | query_key_value | - |
2023-12-18 22:29:45 +08:00
| [ChatGLM3 ](https://huggingface.co/THUDM/chatglm3-6b ) | 6B | query_key_value | chatglm3 |
| [Falcon ](https://huggingface.co/tiiuae ) | 7B/40B/180B | query_key_value | falcon |
| [InternLM ](https://huggingface.co/internlm ) | 7B/20B | q_proj,v_proj | intern |
2023-11-06 17:18:12 +08:00
| [LLaMA ](https://github.com/facebookresearch/llama ) | 7B/13B/33B/65B | q_proj,v_proj | - |
| [LLaMA-2 ](https://huggingface.co/meta-llama ) | 7B/13B/70B | q_proj,v_proj | llama2 |
2023-11-06 12:25:47 +08:00
| [Mistral ](https://huggingface.co/mistralai ) | 7B | q_proj,v_proj | mistral |
2023-12-12 11:39:04 +08:00
| [Mixtral ](https://huggingface.co/mistralai ) | 8x7B | q_proj,v_proj | mistral |
2023-12-18 15:46:45 +08:00
| [Phi-1.5/2 ](https://huggingface.co/microsoft ) | 1.3B/2.7B | Wqkv | - |
2023-12-18 22:29:45 +08:00
| [Qwen ](https://huggingface.co/Qwen ) | 1.8B/7B/14B/72B | c_attn | qwen |
| [XVERSE ](https://huggingface.co/xverse ) | 7B/13B/65B | q_proj,v_proj | xverse |
| [Yi ](https://huggingface.co/01-ai ) | 6B/34B | q_proj,v_proj | yi |
2023-12-29 13:50:24 +08:00
| [Yuan ](https://huggingface.co/IEITYuan ) | 2B/51B/102B | q_proj,v_proj | yuan |
2023-08-07 15:02:02 +08:00
2023-09-10 21:01:20 +08:00
> [!NOTE]
2023-09-10 20:43:56 +08:00
> **Default module** is used for the `--lora_target` argument, you can use `--lora_target all` to specify all the available modules.
>
2023-10-13 13:53:43 +08:00
> For the "base" models, the `--template` argument can be chosen from `default`, `alpaca`, `vicuna` etc. But make sure to use the **corresponding template** for the "chat" models.
2023-10-27 22:15:25 +08:00
2023-11-15 18:04:37 +08:00
Please refer to [constants.py ](src/llmtuner/extras/constants.py ) for a full list of models we supported.
2023-05-31 16:54:06 +08:00
2023-05-31 16:57:43 +08:00
## Supported Training Approaches
2023-05-31 16:54:06 +08:00
2023-08-17 11:00:22 +08:00
| Approach | Full-parameter | Partial-parameter | LoRA | QLoRA |
| ---------------------- | ------------------ | ------------------ | ------------------ | ------------------ |
| Pre-Training | :white_check_mark: | :white_check_mark: | :white_check_mark: | :white_check_mark: |
| Supervised Fine-Tuning | :white_check_mark: | :white_check_mark: | :white_check_mark: | :white_check_mark: |
2023-11-16 02:08:04 +08:00
| Reward Modeling | :white_check_mark: | :white_check_mark: | :white_check_mark: | :white_check_mark: |
| PPO Training | :white_check_mark: | :white_check_mark: | :white_check_mark: | :white_check_mark: |
| DPO Training | :white_check_mark: | :white_check_mark: | :white_check_mark: | :white_check_mark: |
2023-05-31 16:54:06 +08:00
2023-09-10 21:01:20 +08:00
> [!NOTE]
2023-12-18 15:46:45 +08:00
> Use `--quantization_bit 4` argument to enable QLoRA.
2023-08-12 21:23:05 +08:00
2023-05-31 16:54:06 +08:00
## Provided Datasets
2023-11-02 23:10:04 +08:00
< details > < summary > Pre-training datasets< / summary >
- [Wiki Demo (en) ](data/wiki_demo.txt )
- [RefinedWeb (en) ](https://huggingface.co/datasets/tiiuae/falcon-refinedweb )
- [RedPajama V2 (en) ](https://huggingface.co/datasets/togethercomputer/RedPajama-Data-V2 )
- [Wikipedia (en) ](https://huggingface.co/datasets/olm/olm-wikipedia-20221220 )
- [Wikipedia (zh) ](https://huggingface.co/datasets/pleisto/wikipedia-cn-20230720-filtered )
- [Pile (en) ](https://huggingface.co/datasets/EleutherAI/pile )
- [SkyPile (zh) ](https://huggingface.co/datasets/Skywork/SkyPile-150B )
- [The Stack (en) ](https://huggingface.co/datasets/bigcode/the-stack )
- [StarCoder (en) ](https://huggingface.co/datasets/bigcode/starcoderdata )
< / details >
< details > < summary > Supervised fine-tuning datasets< / summary >
- [Stanford Alpaca (en) ](https://github.com/tatsu-lab/stanford_alpaca )
- [Stanford Alpaca (zh) ](https://github.com/ymcui/Chinese-LLaMA-Alpaca )
- [GPT-4 Generated Data (en&zh) ](https://github.com/Instruction-Tuning-with-GPT-4/GPT-4-LLM )
- [Self-cognition (zh) ](data/self_cognition.json )
- [Open Assistant (multilingual) ](https://huggingface.co/datasets/OpenAssistant/oasst1 )
- [ShareGPT (zh) ](https://huggingface.co/datasets/QingyiSi/Alpaca-CoT/tree/main/Chinese-instruction-collection )
- [Guanaco Dataset (multilingual) ](https://huggingface.co/datasets/JosephusCheung/GuanacoDataset )
- [BELLE 2M (zh) ](https://huggingface.co/datasets/BelleGroup/train_2M_CN )
- [BELLE 1M (zh) ](https://huggingface.co/datasets/BelleGroup/train_1M_CN )
- [BELLE 0.5M (zh) ](https://huggingface.co/datasets/BelleGroup/train_0.5M_CN )
- [BELLE Dialogue 0.4M (zh) ](https://huggingface.co/datasets/BelleGroup/generated_chat_0.4M )
- [BELLE School Math 0.25M (zh) ](https://huggingface.co/datasets/BelleGroup/school_math_0.25M )
- [BELLE Multiturn Chat 0.8M (zh) ](https://huggingface.co/datasets/BelleGroup/multiturn_chat_0.8M )
- [UltraChat (en) ](https://github.com/thunlp/UltraChat )
- [LIMA (en) ](https://huggingface.co/datasets/GAIR/lima )
- [OpenPlatypus (en) ](https://huggingface.co/datasets/garage-bAInd/Open-Platypus )
- [CodeAlpaca 20k (en) ](https://huggingface.co/datasets/sahil2801/CodeAlpaca-20k )
- [Alpaca CoT (multilingual) ](https://huggingface.co/datasets/QingyiSi/Alpaca-CoT )
2023-11-15 18:04:37 +08:00
- [OpenOrca (en) ](https://huggingface.co/datasets/Open-Orca/OpenOrca )
2023-11-02 23:10:04 +08:00
- [MathInstruct (en) ](https://huggingface.co/datasets/TIGER-Lab/MathInstruct )
- [Firefly 1.1M (zh) ](https://huggingface.co/datasets/YeungNLP/firefly-train-1.1M )
- [Web QA (zh) ](https://huggingface.co/datasets/suolyer/webqa )
- [WebNovel (zh) ](https://huggingface.co/datasets/zxbsmk/webnovel_cn )
2023-12-01 15:34:50 +08:00
- [Nectar (en) ](https://huggingface.co/datasets/berkeley-nest/Nectar )
2023-12-25 18:29:34 +08:00
- [deepctrl (en&zh) ](https://www.modelscope.cn/datasets/deepctrl/deepctrl-sft-data )
2023-11-02 23:10:04 +08:00
- [Ad Gen (zh) ](https://huggingface.co/datasets/HasturOfficial/adgen )
- [ShareGPT Hyperfiltered (en) ](https://huggingface.co/datasets/totally-not-an-llm/sharegpt-hyperfiltered-3k )
- [ShareGPT4 (en&zh) ](https://huggingface.co/datasets/shibing624/sharegpt_gpt4 )
- [UltraChat 200k (en) ](https://huggingface.co/datasets/HuggingFaceH4/ultrachat_200k )
- [AgentInstruct (en) ](https://huggingface.co/datasets/THUDM/AgentInstruct )
2023-11-02 23:42:49 +08:00
- [LMSYS Chat 1M (en) ](https://huggingface.co/datasets/lmsys/lmsys-chat-1m )
2023-11-02 23:10:04 +08:00
- [Evol Instruct V2 (en) ](https://huggingface.co/datasets/WizardLM/WizardLM_evol_instruct_V2_196k )
< / details >
< details > < summary > Preference datasets< / summary >
- [HH-RLHF (en) ](https://huggingface.co/datasets/Anthropic/hh-rlhf )
- [Open Assistant (multilingual) ](https://huggingface.co/datasets/OpenAssistant/oasst1 )
- [GPT-4 Generated Data (en&zh) ](https://github.com/Instruction-Tuning-with-GPT-4/GPT-4-LLM )
2023-12-01 15:34:50 +08:00
- [Nectar (en) ](https://huggingface.co/datasets/berkeley-nest/Nectar )
2023-11-02 23:10:04 +08:00
< / details >
2023-05-31 16:54:06 +08:00
Please refer to [data/README.md ](data/README.md ) for details.
2023-07-19 20:59:15 +08:00
Some datasets require confirmation before using them, so we recommend logging in with your Hugging Face account using these commands.
2023-05-31 16:54:06 +08:00
```bash
pip install --upgrade huggingface_hub
huggingface-cli login
```
2023-05-29 21:53:02 +08:00
## Requirement
2023-05-31 16:54:06 +08:00
- Python 3.8+ and PyTorch 1.13.1+
2023-05-29 21:53:02 +08:00
- 🤗Transformers, Datasets, Accelerate, PEFT and TRL
2023-09-11 17:31:34 +08:00
- sentencepiece, protobuf and tiktoken
2023-11-13 22:20:35 +08:00
- jieba, rouge-chinese and nltk (used at evaluation and predict)
2023-11-02 23:10:04 +08:00
- gradio and matplotlib (used in web UI)
- uvicorn, fastapi and sse-starlette (used in API)
2023-05-29 21:53:02 +08:00
2023-11-29 12:05:03 +08:00
### Hardware Requirement
2023-12-12 11:39:04 +08:00
| Method | Bits | 7B | 13B | 30B | 65B | 8x7B |
| ------ | ---- | ----- | ----- | ----- | ------ | ------ |
2023-12-23 15:24:27 +08:00
| Full | 16 | 160GB | 320GB | 600GB | 1200GB | 900GB |
2023-12-12 11:39:04 +08:00
| Freeze | 16 | 20GB | 40GB | 120GB | 240GB | 200GB |
| LoRA | 16 | 16GB | 32GB | 80GB | 160GB | 120GB |
| QLoRA | 8 | 10GB | 16GB | 40GB | 80GB | 80GB |
| QLoRA | 4 | 6GB | 12GB | 24GB | 48GB | 32GB |
2023-05-29 21:53:02 +08:00
## Getting Started
### Data Preparation (optional)
2023-11-02 23:10:04 +08:00
Please refer to [data/README.md ](data/README.md ) for checking the details about the format of dataset files. You can either use a single `.json` file or a [dataset loading script ](https://huggingface.co/docs/datasets/dataset_script ) with multiple files to create a custom dataset.
2023-05-29 21:53:02 +08:00
2023-09-10 21:01:20 +08:00
> [!NOTE]
2023-09-10 20:43:56 +08:00
> Please update `data/dataset_info.json` to use your custom dataset. About the format of this file, please refer to `data/README.md`.
2023-05-29 21:53:02 +08:00
### Dependence Installation (optional)
```bash
2023-10-12 21:42:29 +08:00
git clone https://github.com/hiyouga/LLaMA-Factory.git
conda create -n llama_factory python=3.10
conda activate llama_factory
cd LLaMA-Factory
2023-05-29 21:53:02 +08:00
pip install -r requirements.txt
```
2023-07-22 14:29:22 +08:00
If you want to enable the quantized LoRA (QLoRA) on the Windows platform, you will be required to install a pre-built version of `bitsandbytes` library, which supports CUDA 11.1 to 12.1.
```bash
pip install https://github.com/jllllll/bitsandbytes-windows-webui/releases/download/wheels/bitsandbytes-0.39.1-py3-none-win_amd64.whl
```
2023-12-12 19:45:59 +08:00
### Use ModelScope Hub (optional)
2023-12-01 16:11:30 +08:00
2023-12-12 19:45:59 +08:00
If you have trouble with downloading models and datasets from Hugging Face, you can use LLaMA-Factory together with ModelScope in the following manner.
2023-12-01 16:11:30 +08:00
2023-12-01 22:53:15 +08:00
```bash
export USE_MODELSCOPE_HUB=1 # `set USE_MODELSCOPE_HUB=1` for Windows
2023-12-01 16:11:30 +08:00
```
2023-12-01 22:53:15 +08:00
Then you can train the corresponding model by specifying a model ID of the ModelScope Hub. (find a full list of model IDs at [ModelScope Hub ](https://modelscope.cn/models ))
2023-12-01 16:11:30 +08:00
2023-12-01 22:53:15 +08:00
```bash
2023-12-01 22:58:29 +08:00
CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
2023-12-01 22:53:15 +08:00
--model_name_or_path modelscope/Llama-2-7b-ms \
... # arguments (same as above)
2023-12-01 16:11:30 +08:00
```
2023-12-12 19:45:59 +08:00
LLaMA Board also supports using the models and datasets on the ModelScope Hub.
2023-12-01 16:11:30 +08:00
2023-12-01 22:53:15 +08:00
```bash
CUDA_VISIBLE_DEVICES=0 USE_MODELSCOPE_HUB=1 python src/train_web.py
2023-12-01 16:11:30 +08:00
```
2023-08-18 01:51:55 +08:00
### Train on a single GPU
2023-09-10 21:01:20 +08:00
> [!IMPORTANT]
2023-09-10 20:52:21 +08:00
> If you want to train models on multiple GPUs, please refer to [Distributed Training](#distributed-training).
2023-09-10 20:43:56 +08:00
2023-08-18 01:51:55 +08:00
#### Pre-Training
2023-05-29 21:53:02 +08:00
```bash
2023-07-15 16:54:28 +08:00
CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
--stage pt \
2023-05-29 21:53:02 +08:00
--do_train \
2023-12-15 20:53:11 +08:00
--model_name_or_path path_to_llama_model \
2023-05-29 21:53:02 +08:00
--dataset wiki_demo \
--finetuning_type lora \
2023-08-18 11:43:10 +08:00
--lora_target q_proj,v_proj \
2023-05-29 21:53:02 +08:00
--output_dir path_to_pt_checkpoint \
--overwrite_cache \
--per_device_train_batch_size 4 \
--gradient_accumulation_steps 4 \
--lr_scheduler_type cosine \
--logging_steps 10 \
--save_steps 1000 \
--learning_rate 5e-5 \
--num_train_epochs 3.0 \
--plot_loss \
--fp16
2023-05-28 18:09:04 +08:00
```
2023-08-18 01:51:55 +08:00
#### Supervised Fine-Tuning
2023-05-28 18:09:04 +08:00
```bash
2023-07-15 16:54:28 +08:00
CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
--stage sft \
2023-05-28 18:09:04 +08:00
--do_train \
2023-12-15 20:53:11 +08:00
--model_name_or_path path_to_llama_model \
2023-05-29 21:53:02 +08:00
--dataset alpaca_gpt4_en \
2023-07-31 23:33:00 +08:00
--template default \
2023-05-28 18:09:04 +08:00
--finetuning_type lora \
2023-08-18 11:43:10 +08:00
--lora_target q_proj,v_proj \
2023-05-28 18:09:04 +08:00
--output_dir path_to_sft_checkpoint \
--overwrite_cache \
2023-05-29 21:53:02 +08:00
--per_device_train_batch_size 4 \
--gradient_accumulation_steps 4 \
--lr_scheduler_type cosine \
--logging_steps 10 \
--save_steps 1000 \
--learning_rate 5e-5 \
--num_train_epochs 3.0 \
--plot_loss \
--fp16
```
2023-08-18 01:51:55 +08:00
#### Reward Modeling
2023-05-29 21:53:02 +08:00
```bash
2023-07-15 16:54:28 +08:00
CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
--stage rm \
2023-05-29 21:53:02 +08:00
--do_train \
2023-12-15 20:53:11 +08:00
--model_name_or_path path_to_llama_model \
--adapter_name_or_path path_to_sft_checkpoint \
--create_new_adapter \
2023-05-29 21:53:02 +08:00
--dataset comparison_gpt4_en \
2023-07-31 23:33:00 +08:00
--template default \
2023-05-29 21:53:02 +08:00
--finetuning_type lora \
2023-08-18 11:43:10 +08:00
--lora_target q_proj,v_proj \
2023-05-29 21:53:02 +08:00
--output_dir path_to_rm_checkpoint \
2023-08-11 03:02:53 +08:00
--per_device_train_batch_size 2 \
2023-05-29 21:53:02 +08:00
--gradient_accumulation_steps 4 \
2023-05-28 18:09:04 +08:00
--lr_scheduler_type cosine \
--logging_steps 10 \
2023-05-29 21:53:02 +08:00
--save_steps 1000 \
2023-08-18 11:43:10 +08:00
--learning_rate 1e-6 \
2023-05-28 18:09:04 +08:00
--num_train_epochs 1.0 \
2023-05-29 21:53:02 +08:00
--plot_loss \
2023-05-28 18:09:04 +08:00
--fp16
```
2023-05-29 21:53:02 +08:00
2023-08-18 01:51:55 +08:00
#### PPO Training
2023-05-29 21:53:02 +08:00
```bash
2023-07-15 16:54:28 +08:00
CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
--stage ppo \
2023-05-29 21:53:02 +08:00
--do_train \
2023-12-15 20:53:11 +08:00
--model_name_or_path path_to_llama_model \
--adapter_name_or_path path_to_sft_checkpoint \
--create_new_adapter \
2023-05-29 21:53:02 +08:00
--dataset alpaca_gpt4_en \
2023-07-31 23:33:00 +08:00
--template default \
2023-05-29 21:53:02 +08:00
--finetuning_type lora \
2023-08-18 11:43:10 +08:00
--lora_target q_proj,v_proj \
2023-05-29 21:53:02 +08:00
--reward_model path_to_rm_checkpoint \
--output_dir path_to_ppo_checkpoint \
--per_device_train_batch_size 2 \
--gradient_accumulation_steps 4 \
--lr_scheduler_type cosine \
2023-11-17 16:17:36 +08:00
--top_k 0 \
--top_p 0.9 \
2023-05-29 21:53:02 +08:00
--logging_steps 10 \
--save_steps 1000 \
--learning_rate 1e-5 \
--num_train_epochs 1.0 \
2023-08-18 01:51:55 +08:00
--plot_loss \
--fp16
2023-05-29 21:53:02 +08:00
```
2023-11-20 21:39:15 +08:00
> [!WARNING]
2023-11-20 22:52:11 +08:00
> Use `--per_device_train_batch_size=1` for LLaMA-2 models in fp16 PPO training.
2023-11-20 21:39:15 +08:00
2023-08-18 01:51:55 +08:00
#### DPO Training
2023-08-11 03:02:53 +08:00
```bash
CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
--stage dpo \
--do_train \
2023-12-15 20:53:11 +08:00
--model_name_or_path path_to_llama_model \
--adapter_name_or_path path_to_sft_checkpoint \
--create_new_adapter \
2023-08-11 03:02:53 +08:00
--dataset comparison_gpt4_en \
--template default \
--finetuning_type lora \
2023-08-18 11:43:10 +08:00
--lora_target q_proj,v_proj \
2023-08-11 03:02:53 +08:00
--output_dir path_to_dpo_checkpoint \
--per_device_train_batch_size 2 \
--gradient_accumulation_steps 4 \
--lr_scheduler_type cosine \
--logging_steps 10 \
--save_steps 1000 \
--learning_rate 1e-5 \
--num_train_epochs 1.0 \
--plot_loss \
--fp16
```
2023-05-29 21:53:02 +08:00
### Distributed Training
2023-08-12 21:23:05 +08:00
#### Use Huggingface Accelerate
2023-05-29 21:53:02 +08:00
```bash
accelerate config # configure the environment
2023-07-15 16:54:28 +08:00
accelerate launch src/train_bash.py # arguments (same as above)
2023-05-29 21:53:02 +08:00
```
2023-09-10 20:43:56 +08:00
< details > < summary > Example config for LoRA training< / summary >
2023-06-27 22:50:23 +08:00
```yaml
compute_environment: LOCAL_MACHINE
2023-09-10 20:43:56 +08:00
distributed_type: MULTI_GPU
2023-06-27 22:50:23 +08:00
downcast_bf16: 'no'
2023-09-10 20:43:56 +08:00
gpu_ids: all
2023-06-27 22:50:23 +08:00
machine_rank: 0
main_training_function: main
mixed_precision: fp16
num_machines: 1
num_processes: 4
rdzv_backend: static
same_network: true
tpu_env: []
tpu_use_cluster: false
tpu_use_sudo: false
use_cpu: false
```
< / details >
2023-08-12 21:23:05 +08:00
#### Use DeepSpeed
```bash
deepspeed --num_gpus 8 --master_port=9901 src/train_bash.py \
--deepspeed ds_config.json \
... # arguments (same as above)
```
2023-09-10 20:43:56 +08:00
< details > < summary > Example config for full-parameter training with DeepSpeed ZeRO-2< / summary >
2023-08-12 21:23:05 +08:00
```json
{
2023-09-10 21:01:20 +08:00
"train_batch_size": "auto",
2023-08-12 21:23:05 +08:00
"train_micro_batch_size_per_gpu": "auto",
"gradient_accumulation_steps": "auto",
"gradient_clipping": "auto",
"zero_allow_untested_optimizer": true,
"fp16": {
"enabled": "auto",
"loss_scale": 0,
"initial_scale_power": 16,
"loss_scale_window": 1000,
"hysteresis": 2,
"min_loss_scale": 1
2024-01-09 18:31:27 +08:00
},
2023-08-12 21:23:05 +08:00
"zero_optimization": {
"stage": 2,
"allgather_partitions": true,
"allgather_bucket_size": 5e8,
"reduce_scatter": true,
"reduce_bucket_size": 5e8,
"overlap_comm": false,
"contiguous_gradients": true
}
}
```
< / details >
2023-11-20 21:39:15 +08:00
### Merge LoRA weights and export model
2023-07-20 17:23:16 +08:00
```bash
2023-08-18 01:51:55 +08:00
python src/export_model.py \
2023-08-18 11:43:10 +08:00
--model_name_or_path path_to_llama_model \
2023-12-15 20:53:11 +08:00
--adapter_name_or_path path_to_checkpoint \
2023-07-31 23:33:00 +08:00
--template default \
2023-07-20 17:23:16 +08:00
--finetuning_type lora \
2023-12-28 18:47:19 +08:00
--export_dir path_to_export \
--export_size 2 \
--export_legacy_format False
2023-07-20 17:23:16 +08:00
```
2023-12-04 11:02:29 +08:00
> [!WARNING]
2023-12-15 23:44:50 +08:00
> Merging LoRA weights into a quantized model is not supported.
> [!TIP]
2023-12-18 15:46:45 +08:00
> Use `--export_quantization_bit 4` and `--export_quantization_dataset data/c4_demo.json` to quantize the model after merging the LoRA weights.
2023-12-04 11:02:29 +08:00
2023-07-18 00:18:25 +08:00
### API Demo
```bash
python src/api_demo.py \
2023-08-18 11:43:10 +08:00
--model_name_or_path path_to_llama_model \
2023-12-15 20:53:11 +08:00
--adapter_name_or_path path_to_checkpoint \
2023-07-31 23:33:00 +08:00
--template default \
2023-12-15 20:53:11 +08:00
--finetuning_type lora
2023-07-18 00:18:25 +08:00
```
2023-11-20 21:39:15 +08:00
> [!TIP]
2023-09-10 20:43:56 +08:00
> Visit `http://localhost:8000/docs` for API documentation.
2023-07-18 00:18:25 +08:00
### CLI Demo
2023-05-29 21:53:02 +08:00
```bash
2023-07-18 00:18:25 +08:00
python src/cli_demo.py \
2023-08-18 11:43:10 +08:00
--model_name_or_path path_to_llama_model \
2023-12-15 20:53:11 +08:00
--adapter_name_or_path path_to_checkpoint \
2023-07-31 23:33:00 +08:00
--template default \
2023-12-15 20:53:11 +08:00
--finetuning_type lora
2023-05-29 21:53:02 +08:00
```
2023-07-18 17:21:16 +08:00
### Web Demo
```bash
python src/web_demo.py \
2023-08-18 11:43:10 +08:00
--model_name_or_path path_to_llama_model \
2023-12-15 20:53:11 +08:00
--adapter_name_or_path path_to_checkpoint \
2023-07-31 23:33:00 +08:00
--template default \
2023-12-15 20:53:11 +08:00
--finetuning_type lora
2023-07-18 17:21:16 +08:00
```
2023-09-23 21:10:17 +08:00
### Evaluation
```bash
CUDA_VISIBLE_DEVICES=0 python src/evaluate.py \
--model_name_or_path path_to_llama_model \
2023-12-15 20:53:11 +08:00
--adapter_name_or_path path_to_checkpoint \
2023-09-23 21:10:17 +08:00
--template vanilla \
2023-12-15 20:53:11 +08:00
--finetuning_type lora
2023-09-23 21:10:17 +08:00
--task mmlu \
--split test \
--lang en \
--n_shot 5 \
--batch_size 4
```
### Predict
2023-05-29 21:53:02 +08:00
```bash
2023-08-18 01:51:55 +08:00
CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
--stage sft \
2023-09-23 00:34:17 +08:00
--do_predict \
2023-12-15 20:53:11 +08:00
--model_name_or_path path_to_llama_model \
--adapter_name_or_path path_to_checkpoint \
2023-08-18 01:51:55 +08:00
--dataset alpaca_gpt4_en \
2023-07-31 23:33:00 +08:00
--template default \
2023-07-20 17:23:16 +08:00
--finetuning_type lora \
2023-09-23 00:34:17 +08:00
--output_dir path_to_predict_result \
2023-08-18 01:51:55 +08:00
--per_device_eval_batch_size 8 \
--max_samples 100 \
2023-11-20 21:39:15 +08:00
--predict_with_generate \
--fp16
2023-08-18 01:51:55 +08:00
```
2023-11-20 21:39:15 +08:00
> [!WARNING]
> Use `--per_device_train_batch_size=1` for LLaMA-2 models in fp16 predict.
> [!TIP]
2023-09-23 21:10:17 +08:00
> We recommend using `--per_device_eval_batch_size=1` and `--max_target_length 128` at 4/8-bit predict.
2023-05-29 21:53:02 +08:00
2023-10-29 22:07:13 +08:00
## Projects using LLaMA Factory
- **[StarWhisper](https://github.com/Yu-Yang-Li/StarWhisper)**: A large language model for Astronomy, based on ChatGLM2-6B and Qwen-14B.
- **[DISC-LawLLM](https://github.com/FudanDISC/DISC-LawLLM)**: A large language model specialized in Chinese legal domain, based on Baichuan-13B, is capable of retrieving and reasoning on legal knowledge.
- **[Sunsimiao](https://github.com/thomas-yanxin/Sunsimiao)**: A large language model specialized in Chinese medical domain, based on Baichuan-7B and ChatGLM-6B.
2023-10-29 22:53:47 +08:00
- **[CareGPT](https://github.com/WangRongsheng/CareGPT)**: A series of large language models for Chinese medical domain, based on LLaMA2-7B and Baichuan-13B.
2024-01-11 23:18:29 +08:00
- **[MachineMindset](https://github.com/PKU-YuanGroup/Machine-Mindset/)**: MBTI Personality large language model Project, capable of giving any LLM 16 different personality types based on dataset + training method.
2023-11-20 21:39:15 +08:00
> [!TIP]
2023-11-18 11:09:52 +08:00
> If you have a project that should be incorporated, please contact via email or create a pull request.
2023-05-29 21:53:02 +08:00
## License
2023-05-31 16:54:06 +08:00
This repository is licensed under the [Apache-2.0 License ](LICENSE ).
2023-12-29 13:50:24 +08:00
Please follow the model licenses to use the corresponding model weights: [Baichuan ](https://huggingface.co/baichuan-inc/Baichuan-13B-Base/resolve/main/Community%20License%20for%20Baichuan-13B%20Model.pdf ) / [Baichuan2 ](https://huggingface.co/baichuan-inc/Baichuan2-13B-Chat/resolve/main/Community%20License%20for%20Baichuan2%20Model.pdf ) / [BLOOM ](https://huggingface.co/spaces/bigscience/license ) / [ChatGLM3 ](https://github.com/THUDM/ChatGLM3/blob/main/MODEL_LICENSE ) / [Falcon ](https://huggingface.co/tiiuae/falcon-180B/blob/main/LICENSE.txt ) / [InternLM ](https://github.com/InternLM/InternLM#license ) / [LLaMA ](https://github.com/facebookresearch/llama/blob/main/MODEL_CARD.md ) / [LLaMA-2 ](https://ai.meta.com/llama/license/ ) / [Mistral ](LICENSE ) / [Phi-1.5 ](https://huggingface.co/microsoft/phi-1_5/resolve/main/Research%20License.docx ) / [Qwen ](https://github.com/QwenLM/Qwen/blob/main/LICENSE ) / [XVERSE ](https://github.com/xverse-ai/XVERSE-13B/blob/main/MODEL_LICENSE.pdf ) / [Yi ](https://huggingface.co/01-ai/Yi-6B/blob/main/LICENSE ) / [Yuan ](https://github.com/IEIT-Yuan/Yuan-2.0/blob/main/LICENSE-Yuan )
2023-06-16 00:02:17 +08:00
2023-05-29 21:53:02 +08:00
## Citation
2023-07-07 12:06:28 +08:00
If this work is helpful, please kindly cite as:
2023-05-29 21:53:02 +08:00
```bibtex
2023-10-12 21:42:29 +08:00
@Misc {llama-factory,
title = {LLaMA Factory},
2023-05-29 21:53:02 +08:00
author = {hiyouga},
2023-10-12 21:42:29 +08:00
howpublished = {\url{https://github.com/hiyouga/LLaMA-Factory}},
2023-05-29 21:53:02 +08:00
year = {2023}
}
```
## Acknowledgement
2023-10-09 20:02:50 +08:00
This repo benefits from [PEFT ](https://github.com/huggingface/peft ), [QLoRA ](https://github.com/artidoro/qlora ) and [FastChat ](https://github.com/lm-sys/FastChat ). Thanks for their wonderful works.
2023-06-27 23:56:29 +08:00
## Star History
2023-10-12 21:42:29 +08:00
