SFT Data Generator

This is a tool for generating data for Supervised Fine-Tuning (SFT) using large language models like GPT-3.5. It can read data from CSV or Excel files and generate corresponding output data using OpenAI's API based on a given prompt.

Features

• Supports reading CSV and Excel files
• Uses OpenAI's Chat Completion API to generate data
• Supports customizable prompts and models
• Supports batch generation with customizable batch size
• Supports setting the number of generation epochs, i.e., generating multiple rounds for the entire dataset
• Supports JSON format output
• Can set OpenAI API URL and key via command line arguments or environment variables

Usage

1. Install dependencies:

   pip install -r requirements.txt
   

2. Prepare the input data file (CSV or Excel) and prompt template.

3. Run the command:

   python data_generator.py --file_path <input_file_path> --prompt <prompt_template> --model <model_name> [other optional arguments]
   

   Required arguments:

   • --file_path: Path to the input data file, must be a CSV or Excel file.
   • --prompt: Prompt template used for generating data.
   • --model: Name of the OpenAI model used for generating data, e.g., gpt-3.5-turbo.

   Optional arguments:

   • --output_file: Output file path, defaults to input file path + .output.jsonl.
   • --batch_size: Batch size, i.e., the number of samples per concurrent request, defaults to 1.
   • --generate_epoch: Number of generation epochs, i.e., the number of rounds to generate for the entire dataset, defaults to 1.
   • --openai_base_url: Base URL for the OpenAI API, if not given, will use the environment variable OPENAI_API_BASE.
   • --openai_api_key: OpenAI API key, if not given, will use the environment variable OPENAI_API_KEY.
   • --json_output: Whether to use JSON format output, defaults to False.

4. The generated data will be saved in the specified output file, with each line being a JSON object.

Example

python data_generator.py --file_path data.csv --prompt "Please generate a question-answer pair based on the given data:" --model gpt-3.5-turbo --batch_size 10 --generate_epoch 3

This will read the data.csv file, use the prompt "Please generate a question-answer pair based on the given data:", call the gpt-3.5-turbo model with a batch size of 10, and generate 3 epochs of data. The output will be saved in the data.csv.output.jsonl file.