Minbyul/AgentMercury-corpus-all

AgentMercury — full corpus (all tasks, all worlds)

The complete synthesized task set behind AgentMercury: Your Agent Can Synthesize Verifiable Environments for Business Scenarios at Scale: 2,172,500 MCP tool-use investigation tasks instantiated over 4,326 executable worlds.

The paper trains and evaluates on a 10-seed-per-world slice of this corpus (43,300 tasks); this repository is the whole thing — every seed of every world.

	Tasks	Worlds
train	1,940,000	3,865
test (held out)	232,500	461
total	2,172,500	4,326

Train/test worlds are disjoint and the split is inherited from the paper's held-out split, so the 4,630-task evaluation set used in the paper is a subset of test here.

• 96 gzipped JSONL shards, 4.2 GB compressed (37.4 GB raw)
• 500 task seeds per source world — 19 env_ids hold 1,000 tasks, 4,307 env_ids hold 500 tasks (the 19 larger ones are the multi-source env_ids below: 500 seeds each from two worlds)

Row schema

Identical to Minbyul/AgentMercury-corpus-sample, plus three fields. Ids are the same across both repos — a task in the sample carries the same id here.

{
  "id":   "shard_4/env_632967fc__seed0",     // <source shard>/<env_id>__seed<N>
  "name": "env_632967fc__seed0",
  "prompt": [{"role": "system", "content": "You are acting as the following persona: …"}],
  "tools": "[{…tool specs…}]",               // JSON string, never empty
  "metadata": {
    "task_id": "env_632967fc__seed0",
    "env_id":  "env_632967fc",                // the world this task came from
    "env_variant": "",                        // non-empty when one env_id has >1 source world
    "split":   "train",                       // train | test
    "ground_truth": {"rubric": "…", "hard_mask": [true, false, …]},
    "assertions": "[{\"type\": \"record_field_equals\", …}]",   // JSON string
    "arena_seed": "{\"persona\": {…}, \"collections\": {…}, …}",  // JSON string
    "mcp_arena_backend": "local",             // in-process; no docker / e2b
    "persona": "…", "persona_role": "…",
    "touched_services": ["crm", "slack"],
    "af": true                                // cross-system action risk
  }
}

tools, metadata.assertions and metadata.arena_seed are JSON strings — parse them. arena_seed + tools are sufficient to run a task: the in-process local arena (training/code/local_arena.py in the code release) rebuilds the MCP services from the seed.

env_variant: 19 env_ids are backed by more than one source world (two generation runs of the same company profile, or two profiles whose slugs fold to the same ASCII form). Tasks from the secondary world carry a non-empty env_variant and an __x<hash> id suffix — treat (env_id, env_variant) as the true world key.

Composition

Services exercised (worlds touching each): slack 3,916, crm 3,198, email 2,748, billing 1,861, ticket 1,186, support 791, project_management 556, ecommerce 155, inventory 103, finance 45.

Industries (top 12 of 2,292 distinct scenario descriptors): commercial banking (225), banking (56), insurance (49), real estate development (39), automobile manufacturing (36), islamic banking (27), steel manufacturing (26), mobile telecommunications (25), steel production (24), petrochemicals (22), general insurance (21), universal banking (21).

Quick start

from datasets import load_dataset
import json

ds  = load_dataset("Minbyul/AgentMercury-corpus-all", split="train", streaming=True)
row = next(iter(ds))
tools = json.loads(row["tools"])
seed  = json.loads(row["metadata"]["arena_seed"])
print(row["metadata"]["env_id"], len(tools), "tools", seed["services"])

Streaming is recommended — the train split is 4+ GB compressed. To pull one world, filter on metadata.env_id; shard files are grouped by the source generation shard, not by world.

Anonymization

Every world is instantiated from a real-company scenario profile (industry, service mix, scale), but the world itself is fictional: entities, employees, customers, records and conversations are all synthesized. Environment identifiers that carried the seed firm's name as provenance are replaced with content-independent ids env_<8 hex>; the mapping is not published.

Verified on this build:

• the source slug appears in 0 of 2,172,500 rows (checked per row, whole-row, before write);
• the four id-bearing fields (id, name, metadata.task_id, arena_seed.task_id) are rewritten together, so no field disagrees;
• content was additionally audited token-by-token against each source slug on the 43,300-task paper slice: 0 full company-name occurrences; the only matches are generic industry descriptors ("this is a fertilizers business"), which are part of the scenario, not an identity.

Known limitations

• Residual real-brand tokens. 15,087 rows (0.69%) name a real brand as a synthetic counterparty or asset — e.g. "Siemens Mobility GmbH Global" as a customer record, "Bosch GWS 18V Angle Grinder" as an inventory item. Counts: siemens 8,508, bosch 2,232, lantmännen 887, caterpillar 849, samsung 646, schneider electric 637, aramco 374, ikea 232. The tasks were composed 2026-06-29, after the upstream lexicon fix (2026-06-02), but the deterministic scrubber that fix shipped carries a curated real-token list for only five template companies, so brand tokens sourced elsewhere survive. They assert fictional facts about real firms and should be scrubbed before any public release. The probe uses a ~90-token brand list, so the count is a lower bound.
• n_expected_effects is 0 on many rows; grading uses assertions + ground_truth.hard_mask.
• Synthetic email addresses split roughly evenly between the reserved .example TLD and invented .com domains; a few of the latter may collide with real registered domains.
• Task text is English only; scenario profiles span many countries.

Provenance

Generated end-to-end by the authors' own pipeline — env-synth (world synthesis) → dive-synth (task + rubric + verifier synthesis). No third-party dataset is a source or a runtime dependency.

Citation

@article{jeong2026agentmercury,
  title  = {AgentMercury: Your Agent Can Synthesize Verifiable Environments for Business Scenarios at Scale},
  author = {Jeong, Minbyul and Yoon, Chanwoong},
  year   = {2026}
}

License: not set. The 43,300-task paper slice is released Apache-2.0 (see the code release and the public sample). This full 2,172,500-task set is offered under separate licensing (the sample card directs buyers to wjdalsquf@gmail.com). Settle the license before flipping this repository to public.