🛠️ ToolGrad-500: A Tool-Use Dataset Generated by ToolGrad (ACL 2026 Findings)

ToolGrad-500 is a synthetic tool-use dataset targeting single-turn parallel function calling scenarios. The dataset is generated using the ToolGrad framework (ACL 2026 Findings), leveraging textual gradients to optimize tool selection and calling.

---

📊 Dataset Structure

This dataset consists of formatted single-turn conversations suitable for standard conversational trainers.

Splits & Size

• train: 600 high-quality agent sessions (500 positive + 100 negative/irrelevance examples)
• test: 110 curated agent sessions for evaluating tool accuracy on data generated from another LLM

Schema & Design Pattern

Unlike standard multi-turn conversational datasets, ToolGrad-500 is specifically structured for single-turn parallel function calling.

Each conversation in the dataset is exactly one turn consisting of 3 messages:

1. system: The system prompt containing strict format instructions and complete documentation for all available API classes/methods.
2. user: A user query.
3. assistant: A Python-braced list of composed tool invocations to satisfy all parts of the user's request simultaneously.

Role	Description
system	Full API library documentation and strict formatting guidelines.
user	A user query.
assistant	Clean list of concurrent function calls (e.g., [func1(arg=val), func2(arg=val)]).

---

🚀 Quickstart: Loading the Dataset

You can load and inspect this dataset in Python with a single line of code:

from datasets import load_dataset

# Load the ToolGrad-500 dataset
dataset = load_dataset("zhongyi-zhou/toolgrad-500")

# Inspect a sample training conversation
print(dataset["train"][0]["messages"])

---

🔍 Data Sample Preview

Here is a real conversation sample from the dataset demonstrating a single-turn, natural-language user instruction requiring the execution of multiple parallel tools concurrently (taken from line 92 of the training set):

[
  {
    "role": "system",
    "content": "You are an expert in composing functions. You are given a question and a set of possible functions. Based on the question, you will need to make one or more function/tool calls to achieve the purpose...\n\nIf you decide to invoke any of the function(s), you MUST put it in the format of [func_name1(params_name1=params_value1...), func_name2(params)]\nYou SHOULD NOT include any other text in the response.\n\n[... Complete API Documentation Library ...]"
  },
  {
    "role": "user",
    "content": "I'm looking for some new music. Can you show me some popular pop songs, what's trending on YouTube in the US for music, and also give me some information about K-pop girl groups, maybe a random one?"
  },
  {
    "role": "assistant",
    "content": "[Social-tokapi_mobile_version-v1_search_music(keyword='pop', offset=0, count=10, sort_type='2', filter_by='1'), Data-youtube_trending-trending(type='music', country='US'), Music-k_pop-random_girl_group()]"
  }
]

---

🧠 Generation Workflow

The dataset was constructed using the following robust pipeline:

1. Seed Session Generation: 500 different seed trajectories running parallel.
2. Base Model: Generated using gemini-2.5-flash-lite for cost-efficient reasoning.
3. Negative/Irrelevance Curation: Incorporates a target ratio of negative samples where no relevant tools are present, forcing the model to decline gracefully or use fallback logic instead of hallucinating tool calls.

---

📜 Citation

If you use this dataset in your research or applications, please cite the original ToolGrad paper:

@misc{zhou2026toolgradefficienttoolusedataset,
      title={ToolGrad: Efficient Tool-use Dataset Generation with Textual "Gradients"}, 
      author={Zhongyi Zhou and Kohei Uehara and Haoyu Zhang and Jingtao Zhou and Lin Gu and Ruofei Du and Zheng Xu and Tatsuya Harada},
      year={2026},
      eprint={2508.04086},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2508.04086}, 
}