Turnslide: Scalable Multi-Turn Data Synthesis by Walking a Finite-State Machine

2026-10-05

AI Summary

This study addresses the scarcity of multi-turn tool-calling data for small models and the high cost and environmental dependencies of existing synthetic approaches. We propose an efficient, automated data synthesis framework based on finite state machines (FSMs). Specifically, this method models APIs as FSMs to govern valid state transitions and optimizes data quality by specifying target distributions. Furthermore, it leverages a single LLM invocation to transform tool sequences into complete dialogue trajectories, thereby eliminating the need for costly simulated environments. Experimental results demonstrate that fine-tuned models achieve 70.7% accuracy, significantly outperforming baselines, while reducing token consumption by 3.6 to 6.6 times. This work effectively unifies high-quality data generation with low computational overhead.

Abstract

Small language models are inexpensive to serve and can run on private infrastructure, but base models are often not good enough at multi-turn tool calling, and fine-tuning them needs per-API data that rarely exists. Existing synthesis methods are too expensive for high-scale fine-tuning, as they often require mock operational environments for different domains and multiple LLM calls per generated conversation turn. We introduce a fully automated, lightweight synthesis framework that models each API as a finite-state machine, representing the system as abstract states that determine when each tool may be called, producing state-valid sequences of tools; sequences are translated into complete examples with a single LLM call. Rather than optimize diversity, we set a target distribution over the number of turns, the tool sequence and task complexity. We measure data quality by fine-tuning SLMs on generated trajectories, showing that our FSM-based generation significantly improves downstream accuracy over an unmutated baseline and, against existing works, reaches 70.7% full accuracy over 63.4% and 53.7% with 3.6-6.6$\times$ fewer tokens.

Problem

• multi-turn tool calling
• data synthesis
• small language models
• fine-tuning
• scalability

Innovation

• Finite-State Machine
• Multi-Turn Data Synthesis
• Tool Calling
• Small Language Models
• Scalable Fine-Tuning