Data Turnstile Framework for Synthetic Function-Calling Data Generation

Data Turnstile generates high-quality synthetic training data for function calling by decomposing multi-turn interactions into constrained, stepwise generation with error-feedback loops. Fine-tuning a Qwen3-0.6B model on this data improves performance on the BFCL single-turn benchmark.