SMB Bench
Collection
The economic value of agents will soon come from running whole businesses, not from finishing tasks. • 3 items • Updated
Request access to the SMB Bench grader bundle
This repository holds the answer key (private world data), the grader, per-run grader evidence and reasoning returned by each provider for the 18 SMB Bench tasks: six businesses, each in a normal quarter and two changed quarters. Requests are reviewed by the ORO team and may take up to 2 business days (approval is manual). Your details are visible to ORO staff.
You agree not to train, fine-tune or distill any model on this data, and not to redistribute these files. The 18 tasks are declared public and seen. Full terms: the "Terms of use" section of this card (also TERMS.md in this repository). Questions: team@oroagents.com.
Log in or Sign Up to review the conditions and access this dataset content.
SMB Bench: grader bundle and run evidence (gated)
By ORO AI (Shardul Bansal, Seth Schilbe, Jarrod Barnes)
SMB Bench measures whether an AI agent can learn from the delayed consequences of its own decisions while it runs a produce
wholesaler's buying desk for a 90-day quarter; the score is profit in dollars. This repository is the
access-requested tier of release smb-bench-v1. The public tier (results and the tool-call trace of every run: 485 scored runs, 486 cells including one failed run, of 9 models
on 18 tasks) is open in oro-agents/smb-bench, which explains the tasks, the score
and the results; collection: https://huggingface.co/collections/oro-agents/smb-bench-6aec4d315223a590f0bc941a; report:
https://oroagents.com/benchmarks/smb-bench
What is here
This is the grade-only gated tier: enough to regrade every scored run to the cent, and nothing that runs the simulation.
| Path | Content |
|---|---|
worlds/<public task id>/ |
the 18 worlds (12 changed quarters, 6 normal quarters) with private data (the answer key) needed to regrade |
references/, reference_worlds/ |
the 18 reference runs (events, journal, manifest) |
evidence/<model>/<task>/r<k>/ |
per-run grader evidence: events, journal, run manifest, verdict, reward (485 scored final cells) |
reasoning/<model>/<task>/r<k>/ |
the reasoning each provider returned: summaries (reasoning_summary or reasoning_text) for the OpenAI, xAI, Meta, Anthropic and Google models; the full reasoning text (reasoning_text) for DeepSeek and GLM, which return it unsummarised |
task_key.json, TASKS.json |
public task id to hidden mechanism and definition; task pins |
code/, requirements.lock, Dockerfile |
pinned grader code (17 Python files) and a clean install |
regrade.py, regrade_references.py, selftest.py |
regrade all runs, regrade references, check that the checker can fail |
VERIFY.md, TERMS.md, LICENSE |
how to verify; terms |
FILE_MANIFEST.json, EVIDENCE_MANIFEST.json, RELEASE.json, DERIVATION.json |
sha256 of every file; version pins; what was changed from source |
Worlds are derived copies
The worlds in worlds/ and reference_worlds/ are derived copies of the worlds the campaign ran, with the world builder's generation
and calibration metadata removed. Each manifest.json keeps only scenario_id and the sha256 of every world file; each
private/truth.json keeps only what the grader reads: scenario_id, opening_lots, demand_response, line_response and
local_market. Everything that determines the graded economics is byte-for-byte unchanged: the nine private schedules
(private/*.parquet) and public/scenario_public.json. Because each run's event log is bound to its world manifest hash, the first
event's scenario_hash, the event hash chain, run_manifest.json and verdict.json were rebound to the derived worlds; no event payload,
amount or decision changed. DERIVATION.json maps the original and derived sha256 of every changed file. Demand in these worlds is
derived from the public FreshRetailNet-50K dataset (Wang et al., 2025; https://huggingface.co/datasets/Dingdong-Inc/FreshRetailNet-50K,
CC BY 4.0); no fitted parameters ship.
Not here, and never released
The simulation engine, the tool server, the world generator and its calibration data. Replay-based checks (action replay, observation replay, memory replay, public-surface audit, pinned-model check) therefore cannot run from this download; ORO runs them on request through the hosted runner (endpoint and client provided with access: email team@oroagents.com; data policy https://lab-api.oroagents.com/v1/data-policy). New held-out tasks will run only through ORO's hosted runner, as the "ORO-verified" tier. All 18 worlds are declared public and seen: their answer key is in this repository.
Terms of use (full text)
These terms cover the contents of the gated repository oro-agents/smb-bench-grader ("the Data"): the grade-only grader code,
the private world data of the 18 tasks (12 leaderboard, 6 control; the answer key), reference runs, per-run grader evidence, the task key, and
the reasoning each model's provider returned. They do not cover the public repository (oro-agents/smb-bench: results and traces), which is CC BY 4.0.
By requesting access you accept:
- No training. You will not use the Data, or anything derived from it, to train, fine-tune, distill, align or evaluate-for-tuning any machine-learning model, nor include it in any pre-training, post-training or retrieval corpus. Evaluating a model on SMB Bench as a benchmark, and regrading published runs, is allowed.
- No redistribution. You will not publish, re-host, mirror or share the Data, or any part of the grader bundle (worlds, private data, task key, reference runs, model reasoning), with anyone who has not accepted these terms through the gate. You may publish aggregate numbers, short excerpts needed to explain a finding, and your own analysis code.
- Model reasoning. This is the reasoning the providers returned: summaries for most providers, the full text for DeepSeek and GLM. It is provided for research on agent behaviour and may not be used to build a competing model or to imitate a provider's models.
- Seen tasks. The 12 leaderboard tasks and 6 control tasks are declared the public set. Their answer key is released to you, so they are seen tasks. Do not claim a held-out result on them. New held-out tasks will run only through ORO's hosted runner.
- Attribution. Cite SMB Bench (the citation in README.md) in work that uses the Data, and credit FreshRetailNet-50K (CC BY 4.0), from which the worlds' walk-in demand is derived.
- Canary. Keep the canary string in every file, and do not remove it from files you share lawfully with other approved users. Do not paste Data into any tool or service that may train on it.
- Your answers. The details you give at the gate (full name, work email, organisation, intended use) are visible to ORO staff and used only to run the gate, report aggregate access numbers, and contact you about corrections to the release. Ask team@oroagents.com to delete them.
- Revocation. ORO may revoke access at any time, for example on breach. Revocation cannot reach copies already made; the terms still apply to them.
- Not released. The simulation engine, the tool server and the world generator are never released, and nothing here licenses you to reconstruct them. Replay-based checks are run by ORO on request through the hosted runner.
- No warranty. The Data is provided as is. ORO may publish errata and new versions.
Questions: team@oroagents.com
Gate mode and how requests are decided
Mode: manual approval. The ORO team reviews each request (the work email should match the stated organisation) and answers within 2 business days. Reviewers of work that uses SMB Bench can ask team@oroagents.com for direct access. ORO can revoke access at any time. Gating is a contract, not encryption: approved users download plain files. The terms, the canary and the access log are what make misuse detectable and actionable.
Verify it yourself
See VERIFY.md. In short, in a clean Python 3.12 environment from requirements.lock:
python -I regrade.py --root . --evidence evidence --results ../smb-bench/data/runs.parquet. The last line must read
REGRADE OK 485/485 cells match to the cent. A modified events file fails.
Citation
@misc{oro2026smbbench,
title = {SMB Bench: measuring AI's ability to learn from delayed consequences},
author = {Bansal, Shardul and Schilbe, Seth and Barnes, Jarrod and {ORO AI}},
year = {2026},
howpublished = {Hugging Face dataset, oro-agents/smb-bench},
note = {Version smb-bench-v1}
}
The task worlds use FreshRetailNet-50K (Wang et al., 2025, arXiv:2505.16319, CC BY 4.0); please also cite it.
Contact: team@oroagents.com
- Downloads last month
- 42