README.md

---
license: apache-2.0
task_categories:
- text-generation
language:
- en
tags:
- agent
size_categories:
- 1K<n<10K
---

# SAP: State-Guided Data Synthesis with Argument Provenance for Multi-Turn Tool Use

**paper:** [SAP: State-Guided Data Synthesis with Argument Provenance for Multi-Turn Tool Use](https://arxiv.org/abs/2609.06124)

## Dataset Overview
This is a high-quality, executor-validated multi-turn tool-use dataset designed for training agentic language models on long-horizon function calling tasks. The dataset focuses on **argument-level cross-turn dependency grounding**, ensuring tool arguments are sourced from verifiable upstream context (initial state, prior tool returns, or user messages) rather than fabricated values.

![image](https://cdn-uploads.huggingface.co/production/uploads/670a8557222579c05ec3005c/73LhGJl6YNl_mWetgM1yN.png)

- Three stages: (1) FSM skeleton synthesis by $\mathcal{A}_{\text{FSM}}$ with provenance tags;
- (2) per-call planning + executor execution by $\mathcal{A}_{\text{plan}} + \varepsilon$ (two-track output: executor args $\theta^{\text{exec}}$ + provenance metadata $\theta^{\text{prov}}$, parallel grouping, per-call retry);
- (3) post-hoc dialogue synthesis by $\mathcal{A}_{\text{msg}}$.

## Basic Statistics
| Item | Details |
|------|---------|
| Total trajectories | **9,644** |
| Task type | Multi-turn tool use / agent function calling |
| Validation | All trajectories verified against live tool executors |
| Supported benchmarks | BFCL v4 Multi-Turn, τ²-bench |

## Reference Training Setup
The dataset was originally used to train a 4B parameter model via pure SFT (no RL):
- Base model: Qwen3-4B-Instruct
- Optimizer: AdamW, learning rate 1e-6
- Batch size: 128, 10 training epochs
- Training framework: verl