angelo-piletti/cataflow-bench

Cataflow Bench

Benchmark tasks for a product-creation agent. A task is what a Swiss online merchant drops into a batch when new stock arrives: an invoice, a packing list, a price list, a brand guide, notes, and photos on product cards. The agent has to turn that into complete products in the shop's catalog, ready for Shopify or Ricardo, asking the merchant only what no file says.

Every task is generated from a curated catalog of a real store and a recipe; a seed decides what the recipe leaves open. The generator withholds facts on purpose (a missing cost, a currency the invoice does not state, a photo on the wrong card) and records in the plan what the right outcome is, so a run can be scored without a human.

Leaderboard

Every versioned benchmark run on these tasks, newest first. Models are named as the run declared them; the reward is the agent repository's scorer, 0 to 1. Details of every run, the exports and the transcripts are published in cataflow-bench-runs.

r101, version 2

Run on 2026-09-27, 10 tasks. The reward is the scorer's, 0 to 1, averaged over the tasks the model completed; cost is what the platform priced the run at, in USD.

model	effort	provider	tasks	mean reward	cost USD	calls	model time
gemini-3.8-flash	medium	gemini	10/10	0.941	12.31	1190	98 min
gpt-6-luna	medium	openai	10/10	0.870	6.26	623	85 min
gpt-5.6-luna	medium	openai	10/10	0.808	4.19	524	67 min
qwen/qwen3.8-27b-fp8	off	vllm	10/10	0.743	12.82	381	116 min
claude-haiku-4-5-20251001	medium	anthropic	10/10	0.715	5.97	572	57 min
gpt-5-nano-medium	—	—	not run	—	—	—	—

Not run, gpt-5-nano-medium: incapable: asks the same question until the sandbox gives up; 0.16 on the three r101 tasks it finished.

Reward per task:

model	aparanjanparis-com	artigianatopasella-it	drachensilber-de	gisellejewelry-ch	lesunja-com	liliflo-ch	osswald-ch	shop-turmkaffee-ch	simons-kaffee-ch	spielzeug24-ch
gemini-3.8-flash	0.97	0.92	0.97	0.96	0.92	0.93	0.93	0.94	0.97	0.89
gpt-6-luna	0.79	0.73	0.98	0.97	0.84	0.80	0.81	0.91	0.92	0.96
gpt-5.6-luna	0.72	0.89	0.91	0.95	0.53	0.79	0.89	0.90	0.87	0.64
qwen/qwen3.8-27b-fp8	0.28	0.87	0.92	0.94	0.69	0.79	0.80	0.71	0.72	0.73
claude-haiku-4-5-20251001	0.61	0.54	0.79	0.71	0.71	0.71	0.77	0.73	0.80	0.77

Tasks, by the content hash the index gives them:

• aparanjanparis-com-r101 e8f8289a8129
• artigianatopasella-it-r101 924caeae5385
• drachensilber-de-r101 94ddcc43be20
• gisellejewelry-ch-r101 0cad84ef5d7a
• lesunja-com-r101 8b522c219785
• liliflo-ch-r101 846028b9d16f
• osswald-ch-r101 5174c5bfc567
• shop-turmkaffee-ch-r101 65444a998a57
• simons-kaffee-ch-r101 44fce376fd89
• spielzeug24-ch-r101 05e36428d2d4

Layout

tasks/<name>/            the task as the sandbox tools import it
  task.json              the manifest: shop, batches, cards, files, hidden answers
  docs/…  cards/<n>/…    the documents and the card photos the manifest names
truth/<name>/
  plan.json              the answer key: every fact, withheld or not, and the expected products
  recipe.json            the recipe the task was drawn from
tasks.jsonl              one row per task, the index the viewer shows
sync.py                  publishes tasks from the agent repository's data/tasks

Give a model tasks/<name>/ and nothing else. truth/ is for scoring.

The index

field	meaning
name	the task, also its directory under tasks/ and truth/
store	the curated catalog the task was drawn from
split	test for every task so far
generator_version	the generator that made it, <version>+<commit>
seed	what the recipe left open was decided by this
batches, cards, documents, expected_products, hidden_answers	counts across the task's batches
difficulty	the hardest batch, as the recipe rates it
answer_languages	the languages the merchant answers in
scenarios	the situations the recipe put in, such as a duplicate photo or an ambiguous price
files, bytes	the inputs under tasks/<name>/, manifest included
content_sha256	a hash of the inputs: a run records it, so the task it saw is known
truth_sha256	a hash of the plan and the recipe

Running a task

The Cataflow backend's sandbox tools import a task directory as a sandbox shop, start a run, answer the agent's questions from the manifest's hidden answers and export what came out; the agent repository's rewards score the export against truth/<name>/plan.json. A run of every task on several models is a benchmark; its results are published separately, in cataflow-bench-runs, keyed by task name and content_sha256.

Provenance

Card photos, shop names, descriptions and product names come from real stores' public catalogs. The documents are synthetic, rendered by the generator from that catalog. Merchant names on invoices, addresses and every price are invented. Nothing here was written by a merchant.

Publishing

sync.py copies tasks from the agent repository next to this one, validates each plan against the agent's schema, moves the plan and recipe under truth/, hashes the inputs and rewrites the index. A task whose inputs changed since it was published is refused unless --force is given, so a published task never changes silently under a run that cites it. The same run rewrites the leaderboard from every versioned benchmark summary the agent repository holds, reading each model's name, effort and provider from the run it declared.

uv sync
uv run python sync.py --dry-run          # what would be published, and why not
uv run python sync.py                    # publish every task with a valid plan
uv run python sync.py --only '*-r101'    # some of them
git add -A && git commit && git push     # the Hub is the origin