MASDAG: A Multi-agent Workflow Scheduling Simulated DAG Dataset

MASDAG has 1,400 synthetic scheduling cases: 14 DAG variants × 100 cases. Each case has a public input and a matching full record with the answer. The two families are Booking (procedural operational tasks) and SGD (task wording drawn from the Schema-Guided Dialogue schemas). SGD supplies vocabulary, not dialogue transcripts or dependency labels.

Core idea: choose a DAG structure, then fill its nodes with tasks whose semantic dependencies realize that structure.

How cases are built

1. Set the DAG shape first: choose the number of agents and a dependency profile, such as a long path, dense layers, or an hourglass barrier.
2. Fill the structure with meaningful tasks: generate requests whose reads, writes, record keys, validation steps, and side effects express the intended dependencies. Booking uses operational templates; SGD uses existing SGD service, intent, and slot descriptions as task vocabulary.
3. Check the graph and add durations: derive or validate direct edges from those task semantics, compute valid parallel execution batches, and assign every agent a reproducible random integer time from 10 to 100. This is a simulated duration weight, not measured runtime. The rule and seed (20260523) are in src/task_times.py.

Files and task

data/Booking/{public,full,coverage}/   7 variants, 100 cases each
data/SGD/{public,full,coverage}/       7 variants, 100 cases each
src/Booking/                         Booking generators
src/SGD/                             SGD generators
src/rebuild.py                       Recreate all cases
src/task_times.py                    Reproducible random durations
src/validate_release.py              Check the release
sources/sgd/                         Source schemas and attribution
metadata/manifest.json               File hashes and case counts

MicroDAG, MesoDAG, and MacroDAG vary size (3–5, 6–20, and 21–50 agents). DeepDAG, DenseDAG, HourglassDAG, and ImplicitDAG stress graph depth, density, join/split barriers, and implicit wording. Some SGD filenames retain an SGD_ prefix; the enclosing folder identifies the family. coverage files summarize generated graph properties, not model scores.

Each public row contains case_id and public_input.agent_requests[] with agent_id, instruction, and time. Every node in both public and full has the same time. A full row adds ground_truth.dependency_edges, execution_order, execution_batches, transitive_dependency_paths, and parallel_pairs. expected and oracle duplicate selected gold fields for evaluator compatibility.

Task: given a shuffled public request list, predict execution_order: a mapping from agent ID to a 1-based parallel batch. For every edge before → after, the first batch must be smaller than the second. Use only public rows as model input. Match answers by family + variant + case_id, because case IDs repeat across variants.

{"case_id":"case_000000","execution_order":{"agent_000":1,"agent_001":2}}

All Hub configurations use a test split for evaluation; the release has no official train/dev split. For example:

from datasets import load_dataset

public = load_dataset("<namespace>/<dataset-name>", "booking_MicroDAG_public", split="test")
gold = load_dataset("<namespace>/<dataset-name>", "booking_MicroDAG_full", split="test")

Reproduce and verify

Requires Python 3.10+ and the standard library:

python src/validate_release.py
python src/rebuild.py --output-dir ../masdag_rebuilt --verify

The cases and labels are generated, not human annotated real workflows. This repository marks the original MASDAG content Apache-2.0; the included SGD schemas retain their CC BY-SA 4.0 terms and attribution.