Davinci-ToolData

📄 Paper · 💻 Code · 🤗 Davinci-Tool-30B · 🤗 Davinci-Tool-30B-SFT · 🤗 Davinci-ToolData

This repository releases a curated, high-quality subset of Davinci-ToolData, the execution-grounded tool-use corpus used to train Davinci-Tool-30B. It contains 15,166 verified multi-turn trajectories that invoke 4,285 distinct tools, drawn from both data engines: EGTS demonstrations replayed and verified in independently authored environments, and EES trajectories in synthesized single-domain and cross-domain environments. Every trajectory was accepted only after its outcome was checked against executable evidence: a recovered state transition, grounded answer evidence, or executable task predicates.

Performance on τ²-bench, BFCL v4 multi-turn, and MCP-Atlas. Davinci-Tool-30B (after DPO) is hatched; its SFT checkpoint is shown for comparison.

Method

Construction of Davinci-ToolData and training of Davinci-Tool.

Davinci-ToolData is built by two complementary pipelines that share one execution and verification interface:

• EGTS — Execution-Grounded Task Synthesis (authored systems). An explorer interacts with a resettable executable world; user tasks are synthesized from the recorded outcome; the world is reset and the task is replayed with a guide, and the replay is accepted only if it recovers the explored outcome. Structural checks, a hard factual judge, and a soft quality judge filter the results.
• EES — Executable Environment Synthesis (synthesized systems). Audited executable domains (191 domains in 19 categories) produce state-grounded tasks with executable verifiers within one domain (EES-Single), and are composed into shared multi-domain worlds (246 compositions) with shared entities, linking fields, cross-domain policies, and coherence constraints (EES-Cross).

Training has two stages: SFT on the verified demonstrations, then decision-level DPO, which reuses restorable states and executable verifiers to fork the SFT policy at decision points, execute alternative actions from the same state, and keep pairs whose verified downstream outcomes differ (1,150 pairs).

Main results

Models below are trained on the full Davinci-ToolData corpus (30,395 trajectories) described in the paper.

Model	τ²-bench Retail	τ²-bench Telecom	τ²-bench Avg.	BFCL v4 MT Avg.	MCP-Atlas
Qwen3-Coder-480B-A35B-Instruct	77.5	39.5	58.5	43.5	28.1
Kimi-K2-Instruct	70.6	65.8	68.2	33.4	59.6
MiniMax-M2.1	81.6	87.0	84.3	71.4	59.6
GLM-4.7	79.8	98.2	89.0	65.4	55.1
GPT-5.2	81.6	98.2	89.9	28.9	61.8
Claude-Sonnet-4.5	86.0	98.0	92.0	55.3	65.2
Qwen3-Coder-30B-A3B-Instruct	68.7	36.0	52.4	30.4	16.9
+ Toucan-1.5M SFT	65.8	89.5	77.7	41.9	22.5
+ Davinci-ToolData SFT (Davinci-Tool-30B-SFT)	73.7	96.5	85.1	60.6	52.8
+ SFT + DPO (Davinci-Tool-30B)	79.8	96.5	88.2	62.8	56.2

All scores are from the authors' evaluations under the official protocols; MCP-Atlas uses the 89 tasks that do not require external API keys. Under the same backbone and SFT hyperparameters, Davinci-ToolData exceeds the public Toucan-1.5M release by 7.4 / 18.7 / 30.3 points.

Transfer to smaller models (SFT on Davinci-ToolData):

Model	τ²-bench Avg.	BFCL v4 MT Avg.
Qwen3-8B → + Davinci-ToolData	32.5 → 74.6	15.4 → 33.9
Qwen3-14B → + Davinci-ToolData	33.4 → 76.8	21.1 → 41.4

Data sources are complementary (SFT on Qwen3-Coder-30B-A3B-Instruct):

EGTS	EES-Single	EES-Cross	τ²-bench	BFCL v4	MCP-Atlas
✓			78.9	57.4	28.1
	✓		82.0	54.5	29.2
		✓	83.4	46.6	47.2
✓	✓		79.4	57.6	31.5
	✓	✓	82.9	55.6	50.6
✓	✓	✓	85.1	60.6	52.8

Dataset structure

File	Trajectories	Distinct invoked tools	Mean tool calls	Mean assistant turns
davinci_tooldata_open.jsonl	15,166	4,285	9.96	10.86

Each trajectory exposes 28.4 tool schemas on average (6,755 distinct declared tools in total), and 2,421 trajectories contain multiple user turns. Trajectories were generated with DeepSeek-V4-Pro.

Format

Each line is one JSON object:

{
  "messages": [
    {"role": "system", "content": "..."},
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": "...",
     "tool_calls": [{"function": {"name": "...", "arguments": "{...}"}}]},
    {"role": "tool", "content": "..."},
    {"role": "assistant", "content": "..."}
  ],
  "tools": [
    {"type": "function",
     "function": {"name": "...", "description": "...", "parameters": "{\"type\":\"object\", ...}"}}
  ]
}

• Roles are system, user, assistant, and tool.
• tool_calls[].function.arguments and tools[].function.parameters are JSON-encoded strings. Decode parameters with json.loads before passing tools to a chat template that expects objects.

Usage

from datasets import load_dataset

ds = load_dataset("SII-Enigma/Davinci-ToolData", split="train")

Reproduction

The construction code (EGTS, EES-Cross) and configs are in the Davinci-Tool repository.

License

License to be determined. Some trajectories are derived from third-party environments that keep their own licenses and terms.

Citation

@misc{yuan2026davincitoolscalingagentictooluse,
      title={Davinci-Tool: Scaling Agentic Tool-Use Supervision with Verifiable Executable Environments}, 
      author={Xiaoyang Yuan and Ji Zhao and Xingyu Hu and Wenhan Wang and Yutao Hu and Qiaosheng Zhang and Pengfei Liu and Yu Qiao and Yi Bin and Wenqi Shao},
      year={2026},
      eprint={2610.16867},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2610.16867}, 
}