ATOD: A Benchmark for Agentic Task-Oriented Dialogue

Paper · Code & canonical release · Dataset schema · License

ATOD is a synthetic English dialogue benchmark for evaluating systems that must track multiple goals over extended interactions. It covers asynchronous execution, interleaved tasks, goal dependencies, and proactive behavior. The accompanying ATOD-Eval framework evaluates goal completion, efficiency, memory recall, proactivity, and response quality.

This is an author-maintained Hugging Face distribution of the public Amazon Science ATOD release. The original data, attribution, and license are preserved. This repository is hosted in Yifei Zhang's personal namespace; the canonical code and generation pipeline remain on GitHub.

At a glance

Property	Description
Language	English
Data type	Synthetic, multi-domain user–system dialogues
Total dialogues	1,000
Medium complexity	428 dialogues
Complex	572 dialogues
Annotation	Goals, dependencies, lifecycle states, and turn-level state changes
Intended use	Research on task-oriented dialogue, memory, planning, and agent evaluation
License	CC BY-NC 4.0

The medium and complex splits are complexity partitions of the benchmark, not training/validation/test partitions. No separate training or validation split is provided. If you adapt models using these data, disclose that use and establish a held-out evaluation protocol.

Quick start

pip install datasets

from datasets import load_dataset

atod = load_dataset("jimmyzhangyf/ATOD")
print(len(atod["medium"]))   # 428
print(len(atod["complex"]))  # 572

dialogue = atod["medium"][0]
for turn in dialogue["turns"]:
    print(f'{turn["speaker"]}: {turn["utterance"]}')

Each row is one complete dialogue, including its goal and turn-level annotations. To load a single partition:

complex_dialogues = load_dataset("jimmyzhangyf/ATOD", split="complex")

For an exact copy of the original JSON objects, use the raw files:

import json
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="jimmyzhangyf/ATOD",
    filename="medium.json",
    repo_type="dataset",
)
with open(path, encoding="utf-8") as f:
    original_dialogues = json.load(f)

The datasets library infers a shared nested schema and may represent missing optional keys as None. The raw JSON files retain the original objects and values. For reproducible experiments, pin the Hugging Face revision you use.

Data structure

Field	Type	Meaning
dialogue_id	string	Stable dialogue identifier
complexity_class	string	medium or complex
dialogue_index	integer	Generation-time index; use dialogue_id for identification
goal_list	list of objects	Goal descriptions, domains, intents, slots, dependencies, and lifecycle annotations
metadata	object	Generation attributes, including goal count, estimated length, and agentic characteristics
turns	list of objects	Ordered utterances with speaker, turn ID, and goal-state annotations

A goal contains fields such as id, domain, intent, content, core_content, slots, slot_values, dependencies, and status. Lifecycle values are not_mentioned, open, pending, completed, failed, and abandoned.

A turn contains turn_id, speaker (USER or SYSTEM), and utterance; its annotations include all_goals and goal_status_changes. A turn here means one speaker utterance, not a user–system exchange. The metadata field estimated_turns is a generation parameter; compute observed length with len(dialogue["turns"]).

See schema.json for the structural specification. Annotations are evaluation targets: avoid exposing future turns or reference goal states to a system being evaluated unless your experimental protocol explicitly calls for them.

Dataset creation

The released pipeline extracts goal structures from the Schema-Guided Dialogue (SGD) dataset, constructs a goal co-occurrence graph, samples trajectories using random walks, and uses an LLM to annotate trajectories, generate dialogues, and label turn-level goal status.

The files here are the fixed public benchmark artifacts, not newly generated data. Regenerating dialogues may produce different outputs because of model availability, model versions, and stochastic behavior. The upstream SGD dataset is not redistributed here. See the generation pipeline and third-party notices.

Validation and provenance

This distribution mirrors GitHub revision a34ba4d83a3e5b246c1e851088de7959fdea2477. Only the two data-file paths were shortened for this distribution:

Hugging Face file	Original GitHub file	Dialogues
medium.json	data/medium/annotated_dialogues.json	428
complex.json	data/complex/annotated_dialogues.json	572

Both JSON files are byte-identical to that public revision. manifest.json records their SHA-256 checksums, sizes, and source revision.

The release validator checks expected counts, globally unique dialogue IDs, required fields, supported status values, goal/dependency consistency, alternating speakers, common credential patterns, and file checksums. These are structural checks, not guarantees of semantic annotation accuracy. The paper's single-author manual audit concerns evaluator decisions; it should not be interpreted as independent multi-annotator validation of the dataset.

Intended use and limitations

ATOD supports research on goal detection and lifecycle tracking, dependency-aware task completion, memory recall, efficiency, proactivity, and dialogue quality. Evaluation code and metric definitions are available in ATOD-Eval.

• Synthetic conversations may not capture the linguistic diversity, ambiguity, or behavior of real users.
• Generation and annotation can inherit errors or biases from the LLM and the source ontology; domain coverage is limited by that ontology.
• LLM-judged scores depend on prompts, model families, and versions. Report these settings alongside results.
• The benchmark is a fixed research artifact, not a continuously updated source of real-world facts.

Dialogues may contain fictional names, locations, contact strings, reservation details, or financial-task language. They are not intended to describe real individuals; do not treat these strings as verified information or act on them. The dataset is not intended for production or safety-critical decisions, inferring personal attributes, or deployment with real users without additional evaluation.

License and attribution

The released ATOD data are distributed under the Creative Commons Attribution-NonCommercial 4.0 International license. Preserve attribution and comply with the full license. Copyright and third-party notices are retained in NOTICE and THIRD_PARTY_LICENSES.md. Upstream materials remain subject to their original licenses.

Citation

If you use ATOD or ATOD-Eval, please cite the associated paper:

@article{zhang2026atod,
  title={ATOD: An Evaluation Framework and Benchmark for Agentic Task-Oriented Dialogue Systems},
  author={Zhang, Yifei and Nayyeri, Hooshang and Khaziev, Rinat and Yilmaz, Emine and Tur, Gokhan and Hakkani-T{\"u}r, Dilek and Thadakamalla, Hari P},
  journal={arXiv preprint arXiv:2601.11854},
  year={2026}
}

For issues concerning the original benchmark or evaluator, use the GitHub issue tracker. For Hugging Face packaging or loading issues, use this repository's Community tab.