angelo-piletti/cataflow-bench-runs

Cataflow Bench Runs

Every run of a model on a Cataflow Bench task, with everything the run left behind: the products it wrote and the platform files a merchant would import, the score with its components, the run's cost and calls, the full transcript of every job, and the pictures it had generated. A benchmark is one run of every task on several models under a version; this dataset keeps them all, and the leaderboard below is drawn from it.

Leaderboard

Newest benchmark first. Models are named as the run declared them; the reward is the agent repository's scorer, 0 to 1. The tasks are those of cataflow-bench at the content hashes listed under each benchmark.

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

runs/<label>/<task>/         one run: model <label> on task <task>
  result.<batch>.json        the export: products, questions, findings, jobs, every model call
  run.<batch>.json           the run summary: status, policy declared, cost, calls
  score.<batch>.json         the scorer's reward, components and notes
  run.log                    the sandbox tools' log of the run, lines over 2000 characters cut
  <batch>-shopify.csv        what Shopify would import
  <batch>-ricardo.csv, .xlsx what Ricardo would import
  transcripts/<job>.jsonl    every message of every job, in order
  images/<image id>.jpg      the pictures the run generated
results.jsonl                one row per run, the index the viewer shows
benchmarks.jsonl             one row per benchmark version
sync.py                      publishes runs from the agent repository

<label> is <benchmark>.<version>-<variant>, so r101.2-qwen-vllm is the qwen-vllm variant in version 2 of benchmark r101. Runs are never rewritten: a later benchmark is a new version with new labels.

The results index

field	meaning
benchmark, version, run_at	which benchmark the run belongs to and when it ran
label, variant	the run's label and the variant it names
provider, model, reasoning_effort, rates	what the worker declared; the rates in USD per million tokens the cost was priced with
task, task_sha256	the task and the content hash cataflow-bench gives it
status, reward, components, notes	how the run ended and how it scored
cost_usd, policy_calls, model_time_ms	what it cost and how long the model took
products, questions	how many it wrote and asked
error	why there is no score, when there is none; not run: … for a variant the benchmark skipped
run_id, jobs	the platform's run id and the number of job transcripts
generated_images, images_published	pictures the run generated, and how many are here
path	the run's directory in this repository
content_sha256	a hash of the export and transcripts as published

Provenance

Transcripts and products are model output: OpenAI, Anthropic and Google models through their APIs, and a Qwen model served with vLLM. Their providers' terms apply to what may be done with that output. Pictures were generated from the tasks' card photos, which come from real stores' public catalogs. The scores are the agent repository's automatic scorer, no human judged them.

Publishing

sync.py reads every versioned benchmark summary in the agent repository next to this one, copies each run's export, fetches its transcripts by run id and its generated pictures through the local viewer, hashes what it published and rewrites both indexes and the leaderboard. A run whose export changed since it was published is refused unless --force is given.

uv sync
uv run python sync.py --dry-run          # what would be published
uv run python sync.py                    # publish every versioned benchmark
uv run python sync.py --benchmark r101.2 # one of them
git add -A && git commit && git push     # the Hub is the origin