ShengbinYue/paraagent-rl

ParaAgent RL

RL tasks and evaluation criteria for ParaAgent: Reinforcing Parallel Acting in Open-World Tool Environments.

Dataset Composition

Source	Training	Validation
ToolBench	8,665	49
Toucan	4,840	42
Simia	2,879	33
Total	16,384	124

Each line in train.jsonl or validation.jsonl contains one task. Interaction trajectories are generated during RL training.

Unified Schema

All three sources use the same six top-level fields:

Field	Type	Description
task_id	string	Unique task identifier.
dataset	string	toolbench, toucan, or simia.
question	string	Task request.
context	string	Input prefix, including service policies; empty when absent.
target_tool_names	list of strings	Reference tools for task evaluation.
expected	object	Evaluation annotations, described below.

Every expected object has all four keys:

Field	Type	Description
task_type	string or null	write or inquiry for Simia; otherwise null.
actions	JSON text or null	Reference write calls and their arguments, supplied for Simia.
outputs	JSON text or null	Expected results, matching rules, and required tool evidence, supplied for Simia.
dependencies	JSON text or null	Task-specific tool-dependency graph.

Decode non-null actions, outputs, and dependencies with json.loads. Null means unavailable; "[]" means an explicitly empty list.

Getting Started

from datasets import load_dataset

data = load_dataset("ShengbinYue/paraagent-rl")
print(len(data["train"]))       # 16384
print(len(data["validation"]))  # 124

task = data["train"][0]
user_input = task["context"] + task["question"]

The converter and default system prompt are maintained in the ParaAgent code repository. With Python 3.10+ and the Hugging Face CLI installed, run from that repository's root:

python -m pip install pyarrow
hf download ShengbinYue/paraagent-rl \
  --repo-type dataset --local-dir data/rl/paraagent-rl
python scripts/data/prepare_rl.py --data-dir data/rl/paraagent-rl

This creates train.parquet and validation.parquet alongside the JSONL files. See the data preparation guide for conversion and prompt options.

Simia Execution Resources

Simia resources include 2,912 task states at state/<task_id>.json, shared Airline/Retail databases in full_db/, and policies in policies/.

RL training also requires the ParaAgent code, matching ToolEnv catalog, retrieval index, caches, dependency graph, and services.

Citation

If you use this dataset in your research, please cite our paper:

@article{yue2026paraagent,
  title={ParaAgent: Reinforcing Parallel Acting in Open-World Tool Environments},
  author={Yue, Shengbin and Wang, Hongru and Wang, Siyuan and Chen, Xiaoxin and Chen, Wei and Wei, Zhongyu},
  journal={arXiv preprint arXiv:2609.33618},
  year={2026}
}