Self-Evolution

[ English | 中文 ]

Artifact Description

This artifact contains all the program code and sample data required for running the Self-Evolution framework. The overall structure of the artifact and detailed information about the files are as follows:

SELF-EVOLUTION       
│  draw_scores.py   // Script for drawing bar charts of experimental results
│  generation_main.py   // Main program
│  LICENSE
│  main-abl-all.py   // Script for drawing line chart of ablation experiment results
│  README.md
│  README_zh.md
│  requirements.txt // Packages required for program execution
│
├─data  // Datasets
│      AIOps_data.json
│      AIOps_eval_data.json
│      toy_data.json
│      toy_eval_data.json
│
└─src
    │  export_model.py  // Model Merge Script
    │  get_ifd_original.py  // IFD score calculation script
    │  self_evolution_generate.py   // Q&A generation script
    │  train_bash.py    // Model Training Script
    │
    └─llmtuner  // Instruction tuning tool

How to Use

Environment Setup

RAM:8GB

VRAM:32GB

Python:3.10.0

CUDA Version:12.1

OS:Ubuntu 24.04 LTS

Install dependencies:pip install -r requirements.txt

Getting Started

Execute the following command in the Self-Evolution project directory:

python generation_main.py \
    --start_id 0 \
    --end_id 3 \
    --source_data_file data/toy_data.json \
    --identify test \
    --model_path <The path of your initial model> \
    --eval_data data/toy_eval_data.json \
    --eval_to_path eval_result/qas_train \
    --mode hard \
    --train_epoch 1 \
    --ifd_k 5 \
    --batch_size 5 \
    --worker_number 1 \
    --infer_gpu 1 \
    --cuda_device 0 \
    --use_vllm

Reproducibility Instructions

Execute the following command in the Self-Evolution project directory:

python generation_main.py \
    --start_id 0 \
    --end_id 8 \
    --source_data_file data/AIOps_data.json \
    --identify reproduce \
    --model_path <The path of your initial model> \
    --eval_data data/AIOps_eval_data.json \
    --eval_to_path eval_result/qas_train \
    --mode hard \
    --train_epoch 1 \
    --ifd_k 2000 \
    --batch_size 2000 \
    --worker_number 1 \
    --infer_gpu 1 \
    --cuda_device 0 \
    --use_vllm

After the program runs, execute the following command in the Self-Evolution project directory:

python draw_scores.py \
    --father_dir ./eval_result/qas_train/reproduce_generations \
    --res_dir exp_res/col_res

After the above command is executed, the bar chart drawn based on the results of this experiment can be obtained in the exp_res/col_res directory.

Execute the following command in the Self-Evolution project directory:

python main-abl-all.py \
    --father_dir ./eval_result/qas_train/reproduce_generations \
    --res_dir exp_res/abl_res

After the above command is executed, the ablation experiment line chart drawn based on the results of this experiment can be obtained in the exp_res/abl_res directory.

Parameter Description

--start_id: Start iteration number, defaults to 0. If continuing iteration, specify the current number of iterations
--end_id: End iteration number, defaults to 8
--source_data_file: Path to the file containing the original knowledge fragments
--identify: A unique identifier for this workflow; if repeated, it will overwrite the previously generated content and can be identified by the amount of data
--eval_data: Path to the evaluation dataset
--eval_to_path: Path to output the evaluation results
--mode: IFD data filtering mode:

• random - Random selection
• easy - Easy questions
• even - Even extraction
• hard - Difficult questions
• all - Select all data without filtering

--train_epoch: The number of epochs for training in each iteration
--ifd_k: Filter out k pieces of data from historical data
--use_vllm: Use the vllm framework to accelerate inference
--batch_size: The size of data chunks
--worker_number: The maximum number of threads
--infer_gpu: The number of GPUs used during the inference phase
--cuda_device: The GPU identifier used during the model training phase

Data Format

Example of the original knowledge fragment dataset:

[
    {
        "instruction":"<Original Knowledge Fragment>"
    },
    {
        "instruction":"<Original Knowledge Fragment>"
    }
]

Example of the model evaluation dataset:

[
    {
        "instruction":"<Original Knowledge Fragment>",
        "input":"<Question>",
        "output":"<Expected Answer>"
    },
    {
        "instruction":"<Original Knowledge Fragment>",
        "input":"<Question>",
        "output":"<Expected Answer>"
    }
]

Evaluation Results

{
    "predict_bleu-4": 10.067049999999998,
    "predict_rouge-1": 32.988690000000005,
    "predict_rouge-2": 8.616989999999998,
    "predict_rouge-l": 20.33882,
    "predict_runtime": 65.3781,
    "predict_samples_per_second": 0.153,
    "predict_steps_per_second": 0.153
}

Acknowledgements

This project benefits from Cherry_LLM and LLaMA-Factory, the project sample data comes from the 2024 CCF International AIOPS Challenge dataset, thanks to all the authors for their contributions.