Datasets:

cklxx
/

laya-browser-suite-c

license: apache-2.0
language:
  - en
tags:
  - web-agent
  - browser-agent
  - benchmark
  - web-navigation
pretty_name: laya-browser Suite C (held-out multi-step real-site tasks)
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: test
        path: tasks.jsonl

Suite C — 27 multi-step browser tasks on held-out real websites

A small live benchmark for web agents: 27 tasks on 27 real sites, 19 of them needing 4 or more actions (search, then filters, sort, pagination, opening the right result). Every task has a programmatic success check on the final page (URL parameters, title, visible text or observed form state), so no LLM judge is needed.

Built to evaluate cklxx/laya-browser: none of these domains appears in any of its training sources (checked at start-up against the training domain list).

Tasks

tasks.jsonl, one row per task:

name               e.g. "wp-cache-commercial-redis"
start_url          where the agent starts
goal               "Search the plugin directory for 'cache', show only commercial plugins, and open the 'Redis Object Cache' plugin."
scripted_steps     the length / shape of a scripted solution ("4: fill, Enter, filter, link")
check_description  what success means ("URL /plugins/redis-cache")
check_python       the exact check: lambda u, t, x, a: ...   (u = final URL, t = title, x = visible text, a = observed actions)
setup              optional pre-run hook (clear_storage: the site's localStorage is cleared first)

Validation

Every task was validated with a model-free scripted solution (observe → pick the action by label → act): a task was kept only if the script passed its check 3/3 on separate runs and the check failed on the start page (doing nothing cannot pass). The judge scrolls the whole final page before checking, so results do not depend on where the agent left the scroll position. Runner and scripts: code/apps/browser_suite_c.py, code/apps/suite_c_scripts.py, code/apps/SUITE_C.md in cklxx/laya-browser.

Results

agent	passed (27 tasks × 2 runs)
laya-browser v17s (322M, one forward pass per step)	2 / 54
laya-browser v19s	11–14 / 54 (two ×2 runs)

results/ has v19s's per-task results and full step traces. Live sites change and load at different speeds: a single run varies by about ±4 tasks, so repeat runs before comparing agents.

License

Apache-2.0 (task definitions and checks). The sites belong to their owners; run the tasks politely (they are ordinary public search / browse flows, no logins, no purchases).