A simple, powerful, minimal codebase to generate synthetic data using OpenAI, MistralAI, AnthropicAI, or offline inference with vLLM

Create a virtual environment and install the following packages:

• pip install --upgrade pip
• pip install --no-cache-dir -r ./requirements.txt

• Change PROVIDER under params.py
• openai, mistral or anthropic

• Change GPT_MODEL, OPENAI_API_KEY and OUTPUT_FILE_PATH under params.py

• Change MISTRALAI_MODEL, MISTRALAI_API_KEY and OUTPUT_FILE_PATH under params.py

• Change ANTHROPICAI_MODEL, ANTHROPICAI_API_KEY and OUTPUT_FILE_PATH under params.py

• Run with python main.py

• Run with sensei_vllm.py

This example generates 100 input-output pairs per iteration by using a local instance of mistralai/Mixtral-8x7B-Instruct-v0.1 for text generation. The script runs an infinite loop and adds samples to an output file after each iteration.

python sensei_vllm.py --model-id mistralai/Mixtral-8x7B-Instruct-v0.1 --backend vllm --tensor-parallel-size 8 --max_len 1024 --dtype float16 --domain lang --outputs ./ --samples_per_iter 100

To use the system prompts for code change domain from lang to code (--domain code) To use a specific group of topics (--topics_group)

• Change the topics in topics.py
• Change the system contexts in system_messages.py
• Change the number of workers in params.py