Local, private-upload staging package; not a cleared public release. There are 900 expanded inputs: nine task configurations, 100 inputs each. DBSCAN, separate historical pilots and original benchmark example collections are excluded. Finance's ten retained pilot queries are already part of its 100-input collection and are not counted again.

Configuration and bilingual task document	Inputs	Maximum LLM calls	Timeout (seconds)
01-financial-analysis	100	40	900
03-court-form-filling	100	40	900
04-exam-block-sequencing	100	60	1800
05-hotel-book-room	100	20	300
06-healthcare-submit-claim	100	25	300
07-content-flagging	100	20	300
08-customer-service	100	35	600
09-pr-functional-risk-review	100	60	1800
10-citation-check	100	40 / 80 / 150	900 / 1800 / 3600

Episode execution budgets

benchmark_config.json is the versioned source of call/time limits. Each row exports its resolved execution_budget with max_llm_calls and timeout_sec. Citation values above are ordered by 1-50, 51-200 and 201+ bibliography entries; select the tier before execution. DBSCAN also has a configured budget but remains excluded from this package.

One call is one model-request attempt, including retries, subagents and simulated users; parallel requests share one counter. Completion on the last allowed call is permitted, but dispatching another is not. The episode timeout includes LLM/network waiting, tools and CPU computation, excluding setup/reset, grading and between-episode skill updates. An unfinished episode that requires another call or reaches its deadline fails as llm_call_limit_exceeded or timeout; stop outstanding work and ignore late results. Tokens are measured, not capped. See the full execution policy. These are our own starting limits, not upstream defaults or measured runtimes. This package exports the policy; a consuming agent runner must enforce it.

Load with standard Hugging Face Datasets

from datasets import load_dataset
ds = load_dataset("YOUR_NAMESPACE/YOUR_DATASET",
                  name="10-citation-check", split="instances")

instances is the only HF split/container for every configuration. It contains all 100 inputs; there is no training/validation/test allocation. For local use, replace the repository ID with this package's directory.

Rows and file bundles

Each Parquet row contains an episode ID, instruction, an exact input-file inventory, archive paths/prefixes, task-document path, reviewer metadata as JSON, reference status, an optional bibliography entry count and the execution budget.

agent.tar.gz: only episode input files, under their original repository paths.
support.tar.gz: shared mounts, initial guidance and available support/notice files.
reviewer.tar.gz: backend states, references, acquisition evidence and review-only files.
inventories/*.json: exact archive member paths, SHA-256 hashes, sizes and executable bits.
provenance/: original collection manifests, indexes, results and specifications.
docs/source/: verbatim task-document snapshots, not a relocated runnable workspace.

Archives retain original repository-relative member paths. The unmodified data_mount.json files resolve when the relevant bundles are unpacked under the same workspace root. Finance's fixed SEC archives and the court blank PDF are included as necessary shared mounts, not as additional episodes.

Only mount one selected episode's input files and its legitimate shared support into an agent environment. Never mount reviewer bundles, Parquet metadata, other episodes, private references or generation tools as agent input. The word private in source paths denotes a visibility boundary, not an access-control feature on a public HF repository.

Offline output evaluation

The offline evaluator is bundled as benchmark_evaluation/. It grades outputs against expanded-case references, not the original benchmarks' fixed answer keys:

python -m pip install -r benchmark_evaluation/requirements.txt
python -B -m benchmark_evaluation --root /path/to/unpacked-workspace \
  --task 01-financial-analysis --episode 001 --submission /path/to/outputs \
  --result /path/to/grading/result.json --reward /path/to/grading/reward.txt

Unpack the task's agent/support/reviewer bundles into the grader workspace, preserving member paths, and provide its benchmark_config.json. Never expose reviewer material or grading results to the agent. Hotel/claim grading requires a trusted runner-owned final-state/call/time record. Other output checks can run without it, but report that budget compliance was not checked. PR reviews require independent adjudication; citation checks require audited labels. Currently neither collection has those labels, so these tasks return ungradable with a null reward, not a fabricated success/failure. The evaluator documents task-specific scope, prose/semantic limitations, and status/exit codes. The package verify command below checks integrity, not task correctness.

This is a file-backed dataset with offline output checks, not a certified agent runtime or a universal semantic correctness oracle. Initial guidance and available benchmark helpers are preserved. Skill-local scripts/, references/, assets/ and notices are included under skills/ and in support archives at their original paths. Native support Python modules and tool schemas are also included. Original files are unchanged: run helpers from their skill directory, and unpack support archives for workspace-relative guidance links. Third-party packages, benchmark runtimes (such as SOPBench's env package) and services still need installation/configuration. SOP tool modules require a controlled backend; private lookup data is not agent support. Citation guidance's optional scripts/generate_schematic.py belongs to a separate skill absent from the checked-out benchmark and is not bundled. Downloads must not execute bundled PR code or TeX.

Labels and source limitations

No fake references were added to the 100 citation inputs. Their author bibliographies are byte-identical, with 7,356 entries total and natural sizes of 9-972 entries. Citation labels remain unknown, not assumed all-real. The source-provider revision and file hashes are pinned; original arXiv source versions and per-file/version redistribution rights are not certified.

Recorded references in other tasks are construction/replay evidence, not uniformly independently certified gold labels or agent success rates. Task-specific limits, licensing notes and generation methods are in the bilingual task documents and source snapshots.

Integrity and upload

Install requirements.txt, then run:

python tools/package_hf_dataset.py verify --output . --load
python tools/package_hf_dataset.py upload --output . --repo-id YOUR_NAMESPACE/YOUR_DATASET

The upload command defaults to a network-free dry run. After reviewing rights, privacy, bundled source terms and reviewer-data exposure, explicitly add:

--execute --rights-reviewed --reviewer-material-reviewed

Execution creates a private HF dataset or requires an existing private dataset. An existing public repository is refused. A private target is not a substitute for permission: do not execute until every included asset may be uploaded. The exporter did not read HF credentials, upload data or certify release rights.

Downloads last month: 40

Total file size:

388 MB