Minbyul/AgentMercury-corpus

AgentMercury — full RL corpus (private)

⚠️ Private repository. This is the complete corpus actually used for the RL runs in AgentMercury — both splits — with the original environment identifiers intact (they carry the seed firm name as provenance). The redacted public slice is Minbyul/AgentMercury-corpus-sample. Do not make this repo public without first scrubbing the id / name / metadata.task_id / arena_seed.task_id slug layer — the row content is already fictional and audited clean, but the identifiers are not.

Contents

File	Config	Rows	Environments	Size
mcp_ufl_hard_10seed_ascii.TRAIN.jsonl	train	38,670	3,865	638 MB
mcp_ufl_hard_10seed_ascii.TEST.jsonl	test	4,630	461	76 MB

• Ten task seeds per environment (2 TRAIN environments carry 20).
• The two splits are disjoint at the environment level, not just the task level: no environment contributes rows to both. A model cannot have seen a TEST world during training under any seed.
• ascii = the ASCII-normalized variant used for training. TRAIN + TEST = 43,300 tasks over 4,326 environments — the full corpus reported in the paper.
• The TEST split also ships with the code release under data/env-synth-260804/mcp_ufl_hard_10seed_ascii.TEST.jsonl (git-lfs); it is here so the two splits stay together.

What each row is

One MCP agentic tool-use task: a persona-driven autonomous investigation over a synthetic enterprise (CRM, email, Slack, billing, ticketing, project management, and domain-specific systems), graded by an LLM-judge rubric plus programmatic assertions over the resulting database state.

Schema, field-by-field, is documented in the public sample card and in data/CORPUS_FORMAT.md of the code release. In brief:

{
  "id": "shard_4/poc_<env>__seed<N>", "name": "...",
  "prompt": [{"role": "system", "content": "..."}],          // single system message, no user turn
  "tools": "<json string: [{name, description, inputSchema}, ...]>",
  "metadata": {
    "task_type": "mcp_laaj", "task_mode": "cud", "task_id": "...",
    "ground_truth": {"rubric": "...", "hard_mask": [bool, ...], "task_mode": "cud"},
    "assertions": "<json string: [{type, collection, locate, field, value, transition}, ...]>",
    "arena_seed": "<json string: {persona, services, collections, emails, slack_messages}>",
    "mcp_arena_backend": "local", "persona": "...", "persona_role": "...",
    "touched_services": [...], "af": bool, "src_task_mode": "investigation"
  }
}

tools, assertions and arena_seed are JSON strings — parse them. arena_seed + tools are sufficient to reconstruct the environment in-process (training/code/local_arena.py): no docker, no e2b, no external tool server, fresh arena per rollout.

TRAIN split composition

Rows	38,670
Environments	3,865 (10 seeds each)
Shards	48
task_mode / task_type	cud / mcp_laaj (100%)
Cross-system risk (af)	23,184 true / 15,486 false (60.0%)
Top services	slack 33,733 · crm 28,525 · email 24,500 · billing 6,146 · ticket 5,861 · support 5,238 · project_management 4,986
Long tail	ecommerce, inventory, finance, analytics, hr, energy_ops, document_management, fleet_ops, commodities_trading, construction_project, telecom_ops, manufacturing, calendar

Corpus-level statistics from the paper (over all 43,300 tasks): 2,287 distinct industries, 842 unique tools, 222 state-table types, 148 service combinations; 10–26 tools per world (mean 16.1), 1–5 services per task (mean 2.9), 2–10 assertions per task (mean 5.4).

What the reported run actually consumed

For the Qwen3.5-4B GRPO run, 3,200 unique task instances received gradients over 200 steps — about 7.4% of the 43,300-task corpus, drawn from a shuffled stream rather than a curated subset. That 7.4% still spans 2,313 environments (53.5%), 1,438 industries (62.9%), 638 tools (75.8%), 137 state tables (61.7%) and 93 service combinations (62.8%). Sampling few tasks does not mean training on few worlds — which is the point of separating world construction from task instantiation.

Usage

huggingface-cli download Minbyul/AgentMercury-corpus --repo-type dataset \
  --local-dir data/env-synth-260804
# equivalently: AGENTMERCURY_DATASET=Minbyul/AgentMercury-corpus bash scripts/download_data.sh

from datasets import load_dataset
train = load_dataset("Minbyul/AgentMercury-corpus", "train")["train"]   # 38,670
test  = load_dataset("Minbyul/AgentMercury-corpus", "test")["train"]    # 4,630 (held out)

(Both require an authorized token.)

Because the repo is private, scripts/download_data.sh needs an authenticated HF_ACCESS_TOKEN with read access; the public reproduction path is the sample repo plus the in-repo TEST split.

Provenance

Built from the raw investigation corpus investigation_corpus_4783_v2 (~7.6 GB, not shipped) via training/code/build_mcp_ufl.py (raw → mcp_ufl JSONL with arena_seed and per-row tool specs) and training/code/build_260718_10seed.py (10-seed/env filter, hard-weighting, TRAIN/TEST split).

Citation

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