WebsiteBench
Website reconstruction benchmark: agents explore a production-level site through a browser, rebuild a full-stack version, and are scored by hidden functional and visual test cases across 70 tasks.
AI-assisted editorial review against a primary source.
What is this resource?
- Resource type
- Benchmark
- Environment
- Web
- Topics
- Evaluation, Web navigation, Long-horizon tasks, UI design
- Evidence retrieved
- 2026-11-06
- Resource ID
- github:tiger-ai-lab/websitebench
What licenses are recorded?
Code, training data, weights, and documents each have their own entry. A blank entry means the license is unknown.
- Code
- Apache-2.0 (repository LICENSE header; not a weights/data license)
- Data
- Not recorded
- Document
- Not recorded
- Weights
- Not recorded
Evaluation protocol
- Focus
- full-stack website reconstruction from browser exploration, without the source code, visual appearance and functional behavior of the rebuilt site
- Setting
- interactive
- Observations
- Not recorded
- Actions / outputs
- harness-defined browser actions through the task's local browser service
- Evaluation
- Hidden test cases per task. A case passes when its functional checks pass and every visual checkpoint's flat-baseline-normalized RGB SSIM reaches 0.99. A task's reward is passed cases divided by cases, weighted equally.
Scale and counting units
- 70 tasks — released tasks source
- 46,916 test cases — hidden test cases summed over the docs/tasks.md table (computed) source
Agents run through selectable Harbor agents, so observations and actions depend on the agent. Each task requests 8 CPUs, 8 GiB of memory and 20 GiB of disk per environment, with up to 16 hours for the agent.
Compare source-reviewed profiles
Recorded specifications
- leaderboard url
- https://website-bench.com/#results
- repository revision
- 78f5558e8197bb91b3bab63428c12b92060ddc6f
- repository tree sha
- a72e8de335c8f531b655eee7d4f9709dec40b366
Explore related topics
Scope and limitations
Read at commit 78f5558e. The README and docs/tasks.md were reviewed; the 70 tasks and 46,916 hidden cases were checked by summing the docs/tasks.md table. The README's paper badge has no link, so no paper relation is recorded. Apache-2.0 is the README's statement for benchmark code, task definitions and verifier. Leaderboard scores were not checked. Nothing was installed or run.
Primary evidence and verification
Use the original publication or publisher page when citing claims about the resource.
- editorial primary review
Retrieved 2026-11-06T06:21:48Z. Supports: name, kind, summary, environments, topics, scope, notes, details.
Content fingerprint
c79cb821975c34552a46f01349edd0869c2f20c9a85d4b7811155d3fb7f7f0b2 - editorial primary review
Retrieved 2026-11-06T06:21:44Z. Supports: details.benchmark_protocol.
Content fingerprint
d98bd00fb45abc63df0fddd6d9f9f5e669396cd53c157e5e06d6de9d103adf09 - license header
Retrieved 2026-11-06T06:22:02Z. Supports: licenses.code.
Content fingerprint
cda5b289b68650d0643c0ded2d43dc797cb7777c4d943f2b03067250de6d3c26
Suggest a correction
Open a source-backed correction. Include this record ID and the original source. Read our editorial policy and coverage boundaries.