DS²-INSTRUCT: Domain-Specific Data Synthesis for Large Language Models Instruction Tuning

A zero-shot framework that generates domain-specific instruction-tuning datasets without human supervision.

Setup

conda env create -f environment.yml
conda activate ds2

Usage

Local vLLM (Qwen 72B)

Run the full pipeline:

bash shell/run_pipeline_qwen72b.sh cfa

OpenAI

export USE_AZURE_OPENAI=1
bash shell/run_pipeline_azure_openai.sh cfa

Supported Tasks

Task	Domain
cfa	Finance
gsm8k	Grade school math
math	Competition math
pubmedqa	Biomedicine
logiqa	Logical reasoning
gpqa	Graduate-level science
medqa	Medicine

Output

The final output is an SFT dataset in output/{task}_sft_dataset.json