Update study dataset (2 files)
Browse filesExported by scripts/figures.mjs on the dashboard's default task selection.
README.md
ADDED
|
@@ -0,0 +1,44 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: cc-by-4.0
|
| 3 |
+
tags: [human-study, web-agents, grounding, hci]
|
| 4 |
+
---
|
| 5 |
+
|
| 6 |
+
# PageGuide user study — task results
|
| 7 |
+
|
| 8 |
+
Every answered task from the PageGuide web study that the analysis is currently counting, one
|
| 9 |
+
row per task, straight from `study_task_results_v2`.
|
| 10 |
+
|
| 11 |
+
Exported 2026-08-14T04:56:50.743Z · 180 rows · 42 participants.
|
| 12 |
+
|
| 13 |
+
## Conditions
|
| 14 |
+
|
| 15 |
+
Each participant sees both arms, interleaved task by task:
|
| 16 |
+
|
| 17 |
+
- `nongrounding` — the agent reports an answer with no evidence attached.
|
| 18 |
+
- `grounding` — the same claims, with citations into the page and saved image crops.
|
| 19 |
+
|
| 20 |
+
## Which tasks are counted
|
| 21 |
+
|
| 22 |
+
| Facet | Rows | Tasks | Selected by |
|
| 23 |
+
|---|---|---|---|
|
| 24 |
+
| Find × Text | 28 | EDU-v1, HARRY-v1 | the dashboard |
|
| 25 |
+
| Find × Visual | 45 | TREE-V1, SVSF-V1, TESLA-V1 | the dashboard |
|
| 26 |
+
| Guide × Text | 49 | gv2-ms9j3200-u0i9nm, gv2-ms9hwloh-15fuvy, gv2-ed05972e-i5fi3b, gv2-ed35d549-ct71ub | the dashboard |
|
| 27 |
+
| Guide × Visual | 58 | gv2-msf1kfhv-o9o8n8, gv2-msf0vpxs-qucehj, gv2-msf0i7ll-hffde9, gv2-msf28nd3-gig8aj | the dashboard |
|
| 28 |
+
|
| 29 |
+
Some tasks are held out of a card — a disputed answer key, a duplicate re-recording, a run whose
|
| 30 |
+
ground truth cannot be scored in full. The table above is exactly what produced this export.
|
| 31 |
+
|
| 32 |
+
## What is not here
|
| 33 |
+
|
| 34 |
+
Participants' free-text notes are omitted, and the session id is replaced by a per-export
|
| 35 |
+
`participant` integer. Both are deliberate: the notes are the only column that can carry
|
| 36 |
+
something about a person rather than about a task.
|
| 37 |
+
|
| 38 |
+
## Scoring
|
| 39 |
+
|
| 40 |
+
`score_*_precision` / `score_*_recall` are the raw rates; the dashboard combines them as F1.
|
| 41 |
+
For Find that is over the passages picked, for Guide the mean of an F1 over the error types
|
| 42 |
+
named and an F1 over the steps blamed, with a correct "no error" scoring in full on a run that
|
| 43 |
+
contains none. A null precision means the participant predicted nothing — zero true positives —
|
| 44 |
+
so its F1 is 0 rather than missing.
|
rows.csv
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|