Quick Start

ChatPLUG is a library that helps researchers and developers train custom models and build chatbots quickly and easily. To get started, you need to install XDPX in development mode, download the necessary checkpoints, and run the command-line interface (CLI). You can then specify the model configuration file and launch it in a pipeline manner to generate a response.

Requirement

First, install XDPX in development mode:

# navigate to the XDPX directory
cd XDPX
# install the XDPX package
pip install -e .

Download Checkpoints

Next, download the necessary checkpoints:

# navigate to the folder where download.sh is located
cd ..
sh download.sh

Running the Command Line Interface (CLI)

Launch the CLI in a pipeline manner (see ChatPipeline) and specify the model configuration file:

# navigate to the XDPX directory
cd XDPX
# launch the CLI and specify the model configuration file
CUDA_VISIBLE_DEVICES=0 x-script fidchat_new chat_pipeline/chatplug_3.7B_sftv2.6.0_instruction.hjson

CLI Usage

Command	Action
query	generate a response
`#exit`	terminate the session
`#show`	show response details
`#new`	create a new session

Training

To train and tune your own model, you can follow a few simple steps. First, download the dataset you wish to use and preprocess the data by converting it from strings to token IDs. You can then run the training script and visualize the training curve. Finally, you can evaluate your model by creating a new configuration file and editing the appropriate fields to reflect the correct paths to your checkpoints and save directory.

Follow these steps to train your own model:

Step 1: Downloading the Dataset from Belle

# navigate to the data directory
cd data/belle
# install git-lfs
git lfs install
# clone the dataset from Hugging Face
git clone https://huggingface.co/datasets/BelleGroup/train_0.5M_CN
# preprocess the dataset
python process_belle_0.5M.py 
# verify that the processed dataset is available
# $ls data/belle 
# train_0.jsonl dev.jsonl ...

2. Preprocessing Data

Convert the data from strings to token IDs.

Input data format

{
    "context": "我有点烦怎么办", 
    "response": "您好，您有烦是很正常的，可以尝试放松自己，多多参加户外活动以及与朋友的交流，这样可以让自己的情绪得到释放。如果您有负面情绪，也可以咨询专业的心理咨询师，以释放压力。", 
    "passages": "bot_profile:你发微博图开心;你沉迷锻炼;你坚持健康饮食;你想让别人觉得你很酷;你也是;;;knowledge: 方式一：找亲朋好友倾诉。把自己的不顺心的事情和自己的亲朋好友诉说之后，可以起到发泄排解的作用，这样自己的心里会好受很多。方式二;;;knowledge: 方式一：找亲朋好友倾诉。把自己的不顺心的事情和自己的亲朋好友诉说之后，可以起到发泄排解的作用，这样自己的心里会好受很多。方式二：找方法解决掉烦心事。如果这个烦心事可以通过的自己的努力去得以解决，那么自己或者找同伴帮忙一起去解决他，方式三：如果暂时无法解决可以选择性的忘记它或者转移自己的注意力，去做自己喜欢的事情，让自己尽快摆脱不好的情绪。以上是小编提供的几种方式，希望更够帮助小伙伴们。;;;knowledge: 每个人都会有每个人的烦恼，我们的生活经历不同，遇见的事情人物也不同，所以会碰到不一样的烦恼。但是有烦恼也是很正常的，只要看我们如何去对待。那么，下面小编为大家;;;knowledge: 觉得很烦的原因有很多种，或是由于工作还有生活的压力导致，有事情没有和朋友和家人倾述，长时间下去有可能会得抑郁。根据你的情况，建议要乐观点，凡事不要钻牛角尖，应当;;;knowledge: 问题描述： 你好，心烦,有时候是人多想了.总是想到一些将来会发生的事情,这对一个人的心理就形成了压力。缓解压力的方法：大声吼1到2分钟将心里的怨气发泄出来；;;;knowledge: 特别烦，首先应注意对情绪进行调节。保持情绪平稳并舒缓心理压力，可以找朋友或家人倾诉心声，将内心不愉快的事情吐露出来，有利于缓解心理压力改善情绪，进而缓解心烦。", 
    "meta_info": {
        "data_type": "dialogue_emotion", 
        "readme": "v1.1.3 emotion数据", 
        "from": "emotion", 
        "from_file": "empathy.jsonl", 
        "line_num": 1496}
}

Process tokens into ids

# preprocess the data using the specified model configuration file
x-prepro chat_pipeline/chatplug_prepro_sft_instruction.hjson
# verify that the processed data is available
# $ls data/dialogue/sft/chatplug/belle_instruction 
# train_0.pt dev.pt

3. Training

3.1 Runing training script.

# start the training process using the specified model configuration file
x-train chat_pipeline/chatplug_3.7B_train_sftv2.6.0_instruction.hjson

Here: global_batch_size = batch_size * update_freq, and batch_size = GPUs * batch_per_GPU.

3.2 Visualize the training curve The training curve plots are in {save_dir}/plots.

3.3 Evaluate using the training checkpoint

create a new configuration file chatplug_3.7B_sft_belle.hjson, from chatplug_3.7B_sftv2.6.0_instruction.hjson:
edit core_chat_save_dir and core_chat_checkpoint fields to reflect the appropriate paths.

core_chat_save_dir: ../checkpoints/sft/chatplug/3.7B/v2.6.0_epoch20_lr1e-4_bs512/
core_chat_checkpoint: ../checkpoints/sft/chatplug/3.7B/v2.6.0_epoch20_lr1e-4_bs512/checkpoint-6000.pt

ChatPLUG provides an easy-to-use library that allows users to build and train custom chatbots quickly and efficiently.