oro-ai/orobench-trajectories

ORO Bench Agent Trajectories

ORO runs an open competition for shopping agents. Participants submit agent implementations, independent validators execute submitted versions in shopping environments, and participants use feedback to improve future submissions. Repeated evaluations test how agents gather evidence, clarify shopper needs, respond to changes, and complete purchases. Submitted versions are frozen for their race so that competitors face the same evaluation cohort. This archive makes recorded behavior and outcomes available for research, comparison, and red teaming.

ORO Bench is the benchmark behind that competition. This release provides a fixed snapshot for studying tool use, agent behavior, outcomes, and failure cases: 90 composed tasks, 10 sampled agent versions, and 1,800 episodes.

Load the data

The main data is Parquet, with named subsets available in the Hugging Face Dataset Viewer and the datasets library. The Harbor archive is an additional download.

from datasets import load_dataset

repo = "oro-ai/orobench-trajectories"
traces = load_dataset(repo, "trajectories", split="sample", streaming=True)
episode = next(iter(traces))
print(episode["task_id"], episode["outcome"], episode["steps"][0])

tasks = load_dataset(repo, "tasks", split="sample")

Subset	Rows	Contents
trajectories (default)	1,800	Environment interaction traces, structured steps, tool calls, observations, agent/task/run joins, and outcomes
inference_trajectories	1,793	Separately captured inference traces, including available model/reasoning information
tasks	90	Instructions, task family, original task and runtime records, including grading material
episodes	1,800	Episode outcomes, reward components, timestamps, and source hashes
runs	20	Successful run records linked to the recorded episodes
agents	10	Sampled agent/version identifiers

Every subset uses the sample split; this release does not define a train/test partition. task_id joins tasks to episodes and trajectories. eval_run_id joins those records to runs.run_id; agent_version_id joins runs/trajectories to agents.

Trajectory steps preserve their recorded order and source, messages, tool calls and observation results. Variable tool arguments and supplementary objects use JSON strings (arguments_json, extra_json, metrics_json). atif_json preserves the complete source ATIF document; json.loads() reconstructs it. In tasks, task and runtime are JSON strings so heterogeneous benchmark fields remain intact. Supplementary nested episode fields are likewise JSON strings. See SCHEMA.md for types and conversion validation.

Collection and sampling

The release covers one frozen benchmark snapshot. All 90 tasks in that snapshot are included. Agent versions were selected independently of score by SHA-256 ranking of the 159 participants, using the published fixed seed. The release includes the 20 successful runs with recorded episodes for those 10 versions. Each selected agent has 180 recorded episodes, and every retained run has an episode for each of the 90 tasks. Run success describes completion of the evaluation run; it does not mean the agent solved every task. Runs without recorded episodes and the missing-slot table are excluded. Sampling provenance

The environment and inference tracks are separate records of different parts of execution; they should not be assumed to align step for step. The cohort is filtered by run status, not task score: all recorded task outcomes from the retained runs are included, including incorrect orders and seven agent-error episodes.

Public task disclosure

This is a public archive of a completed evaluation snapshot. Task records include task-specific answer sets and grading tables, and traces include verifier evidence. The source_split: private_eval value in the task rows is retained source metadata; it does not describe the visibility of this release. All Hugging Face subsets use sample. These published tasks should be treated as disclosed examples, not as an unseen test set.

Task construction and recorded checks

Each composed task combines an opening request with requirements, shopper disclosures, interaction obligations, and a timeline of changes. Requirements specify properties of the ordered item, such as category, attributes, budget, or a shopper-selected priority. Facets specify what the shopper can reveal in response to an eligible question. Timeline transitions can revise requirements or change price and stock. Obligations specify evidence or interaction that must occur before ordering. The task's grading tables encode which catalog candidates satisfy its predicates; numeric metrics determine reward after the correctness gate passes.

In the tasks subset, json.loads(row["task"]) exposes these definitions under situation and grading. Join an episode to its task with task_id, then decode verdict_checks to inspect the recorded checks. Labels such as need, revised, and market:0 are identifiers within a task; read that task's predicate, obligation, and timeline to determine their exact meaning. Unselected choice branches do not apply. A transition-dependent obligation is satisfied without further action when its transition never fires.

Historical reward gate and check pass rates

In this archived snapshot, an episode had to pass deterministic correctness checks: it had to produce an order, satisfy every applicable task requirement and interaction obligation, and pass the stock and catalog-membership checks. Only then could the family-specific metric contribute to the paid reward; a gate failure recorded a paid reward of zero. Of 1,793 completed episodes, 216 passed the gate (12.0%) and 1,577 failed it (88.0%). The exported aggregate_reward field is null for those 1,577 failures and for seven episodes that ended with an agent error before grading. These two cases should be distinguished using outcome, verdict_correct, and the JSON-encoded reward_components, rather than interpreting every null as either a zero or an ungraded episode. Among the 216 eligible episodes, 17 still have a zero paid reward and 199 have a positive paid reward. These persisted rewards are not official leaderboard standings.

The rates below describe this archived snapshot only. Grading has since changed; these rates do not describe the active competition or predict current scores. They are computed from the recorded episode checks. Each denominator counts episodes where the check is recorded. A conditional obligation can pass when its triggering event never fires; these are not rates restricted to triggered events. Absent task-specific checks are not counted as failures. The seven agent-error episodes have no recorded checks. The overall gate is a conjunction of applicable checks, so individual pass rates should not be averaged to obtain the gate pass rate. final_in_stock and in_stock are duplicate views of the stock check.

Recorded check	What it verifies in this snapshot	Passed / recorded	Pass rate
final_in_stock	The ordered item was in stock at order time; mirrors in_stock.	1,793 / 1,793	100.0%
in_stock	Underlying stock check for the ordered item at order time.	1,793 / 1,793	100.0%
no_illegal_side_effects	The ordered product and SKU exist in the task catalog. Despite its name, this is not a general security or tampering check.	1,793 / 1,793	100.0%
obligation:answered:revision	The agent answers the shopper's revision message after that transition and before ordering.	171 / 736	23.2%
obligation:claims	Claims about the order are truthful and supported by observed listing evidence, including the task's required evidence after a market change.	17 / 59	28.8%
obligation:committed	The first successful cart addition satisfies the task's stated candidate table.	1,249 / 1,733	72.1%
obligation:intended	The first successful cart addition satisfies the task's brand-intent candidate table.	116 / 240	48.3%
obligation:noticed:market:0	The agent observes or signals the first market change before ordering and avoids the item if that change invalidates it.	1,730 / 1,733	99.8%
obligation:noticed:market:1	The same behavior for the second market change, where present; the timeline defines whether it changes price or stock.	179 / 179	100.0%
obligation:survey	The agent views or compares at least three distinct source listings allowed by the task's survey table; variants of one listing count once.	179 / 319	56.1%
requirement:attribute	The order meets the task's stated attribute predicate, such as color, size, material, or a product specification.	787 / 1,202	65.5%
requirement:brand	Records a soft brand preference in this snapshot; it is not a hard brand-match test. Brand-related commitment can be checked separately by obligation:intended.	240 / 240	100.0%
requirement:budget	Order currency matches the requirement and its positive price is within the stated maximum.	613 / 615	99.7%
requirement:budget:cut	The order meets a budget introduced by a timeline transition, when that requirement is active.	1,127 / 1,178	95.7%
requirement:category	The order belongs to the task's required product category.	1,735 / 1,793	96.8%
requirement:choice	The shopper's disclosed priority resolves to a valid choice branch.	110 / 1,035	10.6%
requirement:choice:a	The order meets branch A's predicate when A is selected; its attribute or brand condition comes from the task.	33 / 46	71.7%
requirement:choice:b	The order meets branch B's predicate when B is selected; its attribute or brand condition comes from the task.	36 / 64	56.2%
requirement:need	The order meets an attribute requirement associated with a shopper need disclosed through an eligible question.	423 / 818	51.7%
requirement:need2	The order meets a second question-dependent attribute requirement, where the task defines one.	72 / 179	40.2%
requirement:revised	The order meets the attribute or brand requirement added by the task's revision transition, when active.	346 / 731	47.3%
within_budget	All active budget requirements passed, including any revised budget.	1,740 / 1,793	97.0%

Open competition, red teaming, and generalization

Agent developers can study public feedback and archives; benchmark maintainers must test whether high scores reflect the shopper's request across new task combinations. Correctness checks are measurable proxies for that goal and can have blind spots. Red teaming looks for cases where an agent passes a check while failing the request, exploits a tool or grader, or uses task-specific answer knowledge. Those findings motivate changes to task construction, grading, and admission checks. The grading used in this historical snapshot has since changed.

The intended learning target is general behavior: clarify uncertain needs, ground claims in observed evidence, handle revised constraints, and choose a valid purchase. Published task IDs, answer tables, and fixed trajectories are disclosed examples. Research using them should evaluate on independently constructed tasks and account for overlap in catalogs, predicates, and answer mappings; a different task ID alone does not establish independence. New evaluation cohorts, frozen submissions during a race, and tests against adversarial strategies help reduce overfitting. Publishing an archive does not establish that disclosed or equivalent tasks can never be reused, nor that the grader is immune to manipulation.

Competition protections include submitted-code checks and review for task-specific hardcoding, prohibited access to private answer data, and evaluator tampering. These complement grader tests and task renewal. Learning general shopping strategies from public material is useful; memorized answers and manipulation cannot establish general capability. Recorded checks and rewards are historical measurements, not proof of robustness or current competition performance.

Harbor download

harbor/recorded-jobs.zip contains the Harbor 0.23 recorded job, ATIF trajectories, separate inference captures, verifier results, and episode evidence. Download with huggingface_hub.hf_hub_download, extract it, then run:

uvx --from harbor==0.23.0 harbor view <parent-of-extracted-job-directory> --jobs

All 1,800 environment and 1,793 inference trajectories passed Harbor 0.23 validation. These are recorded trajectories; running or independently grading the tasks requires a runtime adapter and the corresponding environment implementation.

Provenance and license

checksums.sha256 covers the release files. Source hashes, sampling provenance, and validation results are under provenance/. Credential-shaped fields, private source-storage locations, and URL authentication components are filtered by the source formatter. See provenance/security-review.json for the scope and limitations of the credential-pattern review.

No additional dataset-wide license is assigned. The software's MIT license does not automatically cover the data; third-party catalog/provider material retains its applicable terms. Public availability should not be interpreted as an additional license grant.