add enterpriseops_gym dataset

datasets/enterpriseops_gym/LICENSE

@@ -0,0 +1,11 @@
+EnterpriseOps-Gym is distributed by ServiceNow Research under the Apache
+License 2.0. See the upstream sources for the authoritative license text:
+
+  - Dataset: https://huggingface.co/datasets/ServiceNow-AI/EnterpriseOps-Gym
+  - Code:    https://github.com/ServiceNow/EnterpriseOps-Gym
+
+Only the EnterpriseOps-Gym task fields (task id, domain, system/user prompts,
+selected tools, SQL verifiers) that appear in the committed ADP samples are
+reused here, under the terms of that Apache 2.0 license. The ADP conversion
+scripts in this directory are covered by the Agent Data Protocol repository
+license.

datasets/enterpriseops_gym/README.md

@@ -0,0 +1,140 @@
+# EnterpriseOps-Gym
+
+## Description
+
+EnterpriseOps-Gym (ServiceNow Research) is a containerized, resettable
+enterprise simulation benchmark for evaluating LLM agents on **stateful,
+multi-step planning and tool use** across realistic enterprise workflows. It
+contains 1,150 expert-curated tasks over 8 domains — **Calendar, CSM, Drive,
+Email, HR, ITSM, Teams, and Hybrid** — backed by 164 relational database tables
+and ~512 MCP tools. Trajectories average ~9 steps (up to 34).
+
+Unlike static, rollout-based datasets, EnterpriseOps-Gym is **evaluated by SQL
+verifiers that check the final environment state**, not by matching an action
+sequence. Each published task carries a system prompt, a user request, the
+oracle tool set, the MCP server configuration, and the SQL verifiers.
+
+ADP stores each agent rollout as an OpenAI-style chat trajectory
+(`system` / `user` / `assistant` / `tool` messages) where assistant turns carry
+MCP tool calls and `tool` turns carry the environment observations. This is the
+same raw shape used by the `crmarena` and `enterpriselab` datasets, so
+EnterpriseOps-Gym reuses ADP's shared `raw_to_atif` / `atif_to_std` converters.
+
+## Dataset Information
+
+- **Original Paper**: [EnterpriseOps-Gym: Environments and Evaluations for Stateful Agentic Planning and Tool Use in Enterprise Settings](https://arxiv.org/abs/2603.13594)
+- **Original Repository**: https://github.com/ServiceNow/EnterpriseOps-Gym
+- **Source Data**: https://huggingface.co/datasets/ServiceNow-AI/EnterpriseOps-Gym
+  (public split ≈ 60% of tasks; configs `oracle`, `plus_5_tools`,
+  `plus_10_tools`, `plus_15_tools`; one split per domain)
+- **Related ADP paper**: https://arxiv.org/abs/2510.24702
+- **License**: Apache 2.0 (see [`LICENSE`](LICENSE) and the upstream sources).
+- **Domains (8)**: `calendar`, `csm`, `drive`, `email`, `hr`, `itsm`, `teams`,
+  `hybrid`.
+- **Tools**: declared in [`metadata.json`](metadata.json). Names are the real
+  EnterpriseOps-Gym MCP tools (taken from each task's `selected_tools`);
+  parameter schemas are inferred from the sample tool calls because the full
+  tool schemas live inside the containerized MCP servers, not the public repo.
+
+## Trajectories are generated, not downloaded
+
+**Important:** The Hugging Face dataset publishes *tasks + SQL verifiers*, **not**
+agent trajectories. A trajectory only exists once an agent is run against the
+live, containerized gym MCP servers. Use
+[`generate_trajectories.py`](generate_trajectories.py):
+
+1. **`--from-hf`** (default, no gym/containers needed) — download the published
+   tasks and build *illustrative reference trajectories* grounded in the real
+   task fields:
+   ```bash
+   export PYTHONPATH=`pwd`:$PYTHONPATH
+   python datasets/enterpriseops_gym/generate_trajectories.py --from-hf \
+       --config oracle --per-domain 1 \
+       --output datasets/enterpriseops_gym/enterpriseops_gym_gold.json
+   ```
+   (Needs the commented extras in `requirements.txt`: `datasets`,
+   `huggingface_hub`, `pandas`, `pyarrow`.)
+
+2. **`--from-results`** — convert result logs from EnterpriseOps-Gym's own
+   `evaluate.py` into authentic gold trajectories:
+   ```bash
+   python datasets/enterpriseops_gym/generate_trajectories.py \
+       --from-results /path/to/EnterpriseOps-Gym/results \
+       --output datasets/enterpriseops_gym/enterpriseops_gym_gold.json
+   ```
+
+3. **`--run`** — placeholder for a live driver against the gym MCP servers
+   (requires the containerized environment + model credentials).
+
+`enterpriseops_gym_gold.json` is a local build artifact (git-ignored). The
+committed `sample_raw.json` is used as the fallback source when it is absent.
+
+### About the committed `sample_raw.json`
+
+Because authentic trajectories require running the containerized gym, the five
+committed samples are **illustrative reference trajectories**: the task id,
+domain, system/user prompts, oracle tool list, and SQL verifiers are the real
+published EnterpriseOps-Gym task, while the tool-call arguments and observations
+are compact, representative values that demonstrate the trajectory shape. Each
+record carries a `note` field. Example access tokens in `gym_servers_config`
+are redacted. Regenerate with `--from-results` + `make_samples.py` to replace
+the illustrative steps with authentic ones.
+
+## Regenerating the committed samples
+
+```bash
+export PYTHONPATH=`pwd`:$PYTHONPATH
+python datasets/enterpriseops_gym/make_samples.py --num-samples 5
+```
+
+`make_samples.py` reads `enterpriseops_gym_gold.json` when present, otherwise
+falls back to the committed `sample_raw.json`. It writes `metadata.json`,
+`sample_raw.json`, `sample_atif.json`, `sample_std.json`, and
+`sample_sft/{openhands_sdk,sweagent}.json`.
+
+## Schema Mapping
+
+- Raw `system` / `user` / `assistant` messages are preserved as ATIF steps.
+- Raw assistant `tool_calls` become ATIF `ToolCall` entries with the original
+  MCP tool names and JSON arguments.
+- Raw `tool` messages become ATIF observations, linked to the preceding
+  assistant tool call.
+- `atif_to_std.py` applies ADP's shared tool-call normalization while retaining
+  EnterpriseOps-Gym-specific tools declared in `metadata.json`.
+- `sample_sft/openhands_sdk.json` is generated from the standardized ATIF sample
+  through the shared OpenHands SDK converter.
+
+Building samples (raw → ATIF → std → SFT) needs `pydantic>=2.12` and the
+OpenHands SDK. Pulling tasks from Hugging Face additionally needs `datasets`,
+`huggingface_hub`, `pandas`, and `pyarrow` (see the commented extras in
+`requirements.txt`).
+
+Diagnose OpenHands import problems with:
+
+```bash
+python datasets/enterpriseops_gym/make_samples.py --check-openhands
+```
+
+## Validation
+
+```bash
+export PYTHONPATH=`pwd`:$PYTHONPATH
+python -m pytest tests/test_dataset_structure.py -v -k enterpriseops_gym
+python -m pytest tests/test_raw_schemas.py -v -k enterpriseops_gym
+python -m pytest tests/test_standardized_schemas.py -v -k enterpriseops_gym
+```
+
+## Citation
+
+Please cite EnterpriseOps-Gym using the citation requested by the upstream
+authors:
+
+```bibtex
+@article{enterpriseopsgym2026,
+  title={EnterpriseOps-Gym: Environments and Evaluations for Stateful Agentic
+         Planning and Tool Use in Enterprise Settings},
+  author={ServiceNow Research},
+  journal={arXiv preprint arXiv:2603.13594},
+  year={2026}
+}
+```

datasets/enterpriseops_gym/atif_to_std.py

@@ -0,0 +1,7 @@
+# ruff: noqa: E402, I001
+
+from scripts.atif_to_std_common import main
+
+
+if __name__ == "__main__":
+    main(__file__)

datasets/enterpriseops_gym/extract_raw.py

@@ -0,0 +1,45 @@
+from __future__ import annotations
+
+import argparse
+import json
+from pathlib import Path
+from typing import Any
+
+DATASET_DIR = Path(__file__).resolve().parent
+# Full trajectories produced by generate_trajectories.py are written here.
+DEFAULT_SOURCE = DATASET_DIR / "enterpriseops_gym_gold.json"
+# Committed reference trajectories used when the full gold file is absent.
+FALLBACK_SOURCE = DATASET_DIR / "sample_raw.json"
+
+
+def load_records(source: Path) -> list[dict[str, Any]]:
+    data = json.loads(source.read_text(encoding="utf-8"))
+    if not isinstance(data, list):
+        raise ValueError(f"Expected {source} to contain a JSON list")
+    return [record for record in data if isinstance(record, dict)]
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description="Extract EnterpriseOps-Gym raw records as JSONL")
+    parser.add_argument(
+        "--source",
+        type=Path,
+        default=None,
+        help=(
+            "Path to EnterpriseOps-Gym trajectories JSON. Defaults to "
+            "enterpriseops_gym_gold.json when present, otherwise falls back to the "
+            "committed sample_raw.json."
+        ),
+    )
+    args = parser.parse_args()
+
+    source = args.source
+    if source is None:
+        source = DEFAULT_SOURCE if DEFAULT_SOURCE.exists() else FALLBACK_SOURCE
+
+    for record in load_records(source):
+        print(json.dumps(record, ensure_ascii=False))
+
+
+if __name__ == "__main__":
+    main()