https://github.com/ssbuild/llm_rlhf
realize the reinforcement learning training for gpt2 llama bloom and so on llm model
https://github.com/ssbuild/llm_rlhf
llm llm-rlhf lora reward rlhf trl trlx
Last synced: 30 days ago
JSON representation
realize the reinforcement learning training for gpt2 llama bloom and so on llm model
- Host: GitHub
- URL: https://github.com/ssbuild/llm_rlhf
- Owner: ssbuild
- Created: 2023-04-19T10:46:13.000Z (about 2 years ago)
- Default Branch: main
- Last Pushed: 2023-09-19T09:03:03.000Z (over 1 year ago)
- Last Synced: 2025-04-18T12:17:54.798Z (about 1 month ago)
- Topics: llm, llm-rlhf, lora, reward, rlhf, trl, trlx
- Language: Python
- Homepage:
- Size: 388 KB
- Stars: 26
- Watchers: 1
- Forks: 2
- Open Issues: 3
-
Metadata Files:
- Readme: README.MD
Awesome Lists containing this project
README
## llm reinforcement learning
Realize the reinforcement learning training for gpt2 llama bloom,cpm-ant and so on.
## update information
- [deep_training](https://github.com/ssbuild/deep_training)
```text
06-13 fix llama resize_token_embeddings
06-01 支持lora deepspeed 训练，0.1.9 和 0.1.10合并
05-27 add qlora transformers>=4.30
```
## install
python >= 3.10
- pip install -U -r requirements.txt
- 如果无法安装，可以切换官方源 pip install -i https://pypi.org/simple -U -r requirements.txt
## weigtht select one is suitable for you
支持且不限于以下权重
- [bloom预训练模型](https://huggingface.co/bigscience)
- [bloom第三方中文训练模型](https://huggingface.co/Langboat/bloom-6b4-zh) # 注意 需要修改tokenizer_config.json BloomTokenizer -> BloomTokenizerFast
- [tigerbot](https://huggingface.co/TigerResearch)
- [opt预训练模型](https://huggingface.co/facebook)
- [llama 官方权重转换](https://huggingface.co/decapoda-research) # llama 词典等下载地址 https://huggingface.co/hf-internal-testing/llama-tokenizer
- [llama vicuna-7B第三方权重1](https://huggingface.co/TheBloke/vicuna-7B-1.1-HF)
- [llama vicuna-7B第三方权重2](https://huggingface.co/Tribbiani/vicuna-7b)
- [cpm-ant-10b](https://huggingface.co/openbmb/cpm-ant-10b)
- [rwkv](https://huggingface.co/BlinkDL) 需要转换权重
- [rwkv](https://huggingface.co/RWKV/rwkv-4-169m-pile) hf 权重
## data sample
```json
{
"prompt": "\n\nHuman:你能描述如何清除笔记本计算机屏幕上的指纹和污垢的步骤吗?\n\nAssistant:是的,也许。为了清洁你的屏幕,你首先需要使用微纤维布或软溼布,轻轻地擦拭屏幕表面。接下来,你会想抓住一个软,无毛巾,微纤维清洁布,轻轻地擦回屏幕上去除指纹和污垢。\n\nHuman:我可以将伊索普罗素酒精喷射到布上,并以这种方式清洗它吗?\n\nAssistant:",
"response": " 是的,你可以这样做,以帮助布从屏幕上收集更多的污垢。 请确保始终使用干净的软布,而不是一块粗糙的、粗糙的或结构化的材料,并确保它是无毛布的。",
"chosen": " 是的,你可以这样做,以帮助布从屏幕上收集更多的污垢。 请确保始终使用干净的软布,而不是一块粗糙的、粗糙的或结构化的材料,并确保它是无毛布的。",
"rejected": " 是的,你可以直接将它喷射到布上。"
}
```
## infer
# infer_finetuning.py 推理微调模型
# infer_lora_finetuning.py 推理微调模型
# infer_ptuning.py 推理p-tuning-v2微调模型
python infer_finetuning.py
## training
```text
#制作数据
python data_utils.py
注: num_process_worker 为多进程制作数据 ， 如果数据量较大 ， 适当调大至cpu数量
dataHelper.make_dataset_with_args(data_args.train_file,mixed_data=False, shuffle=True,mode='train',num_process_worker=0)
#训练
python train.py
```
## 训练参数
[训练参数](args.MD)
## 友情链接
- [pytorch-task-example](https://github.com/ssbuild/pytorch-task-example)
- [chatmoss_finetuning](https://github.com/ssbuild/chatmoss_finetuning)
- [chatglm_finetuning](https://github.com/ssbuild/chatglm_finetuning)
- [chatglm2_finetuning](https://github.com/ssbuild/chatglm2_finetuning)
- [t5_finetuning](https://github.com/ssbuild/t5_finetuning)
- [llm_finetuning](https://github.com/ssbuild/llm_finetuning)
- [llm_rlhf](https://github.com/ssbuild/llm_rlhf)
- [chatglm_rlhf](https://github.com/ssbuild/chatglm_rlhf)
- [t5_rlhf](https://github.com/ssbuild/t5_rlhf)
- [rwkv_finetuning](https://github.com/ssbuild/rwkv_finetuning)
- [baichuan_finetuning](https://github.com/ssbuild/baichuan_finetuning)
- [baichuan2_finetuning](https://github.com/ssbuild/baichuan2_finetuning)
- [internlm_finetuning](https://github.com/ssbuild/internlm_finetuning)
- [qwen_finetuning](https://github.com/ssbuild/qwen_finetuning)
- [xverse_finetuning](https://github.com/ssbuild/xverse_finetuning)
- [aigc_serving](https://github.com/ssbuild/aigc_serving)
##
纯粹而干净的代码