This project is a Python-based tool for generating fine-tuning datasets using the DeepSeek API. It creates single-turn Q&A pairs or multi-turn conversations in various styles, based on customizable topics and categories. It also includes a dataset validator to ensure quality before fine-tuning.

• Generate realistic Q&A examples with customizable styles

• Support for single or multi-turn conversations

• Style options: helpful, corporate, casual, technical, creative, educational

• Validate .jsonl datasets before fine-tuning

• Topic categories covering technology, business, education, lifestyle, and more

• Rate limiting support and environment configuration via .env

• Python 3.7+

• DeepSeek API Key

• Dependencies:

  bash

  pip install -r requirements.txt

1. Clone the repository:

   bash

   git clone https://github.com/luisriverag/deepseek-api_dataset-generator.gitcd deepseek-api_dataset-generator

2. Create a .env file:

   CopyEdit

   python3 generator.py --create-env

3. Edit .env and add your DeepSeek API key:

   env

   DEEPSEEK_API_KEY=your_deepseek_api_key_here

python3 generator.py

python3 generator.py \ --output my_dataset.jsonl \ --count 100 \ --style technical \ --categories technology education \ --conversation-turns 2

python3 generator.py --validate-only --output my_dataset.jsonl

Available topic categories:

• technology

• business

• education

• lifestyle

• creative

• science

• all (default)

Each line in the output .jsonl file follows the format:

json

{ "messages": [ {"role": "user", "content": "What is quantum computing?"}, {"role": "assistant", "content": "Quantum computing uses principles of quantum mechanics to perform computations..."} ]}

Ensures:

• Proper JSON structure

• Valid roles: user, assistant, system

• Token estimation

• Warnings for long/empty content

1. Review your dataset.

2. Use it for fine-tuning a model that supports DeepSeek-style training.

3. Monitor performance and adjust generation parameters as needed.

MIT License. See LICENSE for details.