Thang203's picture
Update study dataset (2 files)
9bdbd63 verified
|
Raw
History Blame Contribute Delete
1.97 kB
metadata
license: cc-by-4.0
tags:
  - human-study
  - web-agents
  - grounding
  - hci

PageGuide user study — task results

Every answered task from the PageGuide web study that the analysis is currently counting, one row per task, straight from study_task_results_v2.

Exported 2026-08-14T05:06:35.500Z · 217 rows · 43 participants.

Conditions

Each participant sees both arms, interleaved task by task:

  • nongrounding — the agent reports an answer with no evidence attached.
  • grounding — the same claims, with citations into the page and saved image crops.

Which tasks are counted

Facet Rows Tasks Selected by
Find × Text 65 EDU-v1, MUFC-V1-TEXT, NVIDA-V1, HARRY-v1 the committed defaults
Find × Visual 45 TREE-V1, SVSF-V1, TESLA-V1 the committed defaults
Guide × Text 49 gv2-ms9j3200-u0i9nm, gv2-ms9hwloh-15fuvy, gv2-ed05972e-i5fi3b, gv2-ed35d549-ct71ub the committed defaults
Guide × Visual 58 gv2-msf1kfhv-o9o8n8, gv2-msf0vpxs-qucehj, gv2-msf0i7ll-hffde9, gv2-msf28nd3-gig8aj the committed defaults

Some tasks are held out of a card — a disputed answer key, a duplicate re-recording, a run whose ground truth cannot be scored in full. The table above is exactly what produced this export.

What is not here

Participants' free-text notes are omitted, and the session id is replaced by a per-export participant integer. Both are deliberate: the notes are the only column that can carry something about a person rather than about a task.

Scoring

score_*_precision / score_*_recall are the raw rates; the dashboard combines them as F1. For Find that is over the passages picked, for Guide the mean of an F1 over the error types named and an F1 over the steps blamed, with a correct "no error" scoring in full on a run that contains none. A null precision means the participant predicted nothing — zero true positives — so its F1 is 0 rather than missing.