Upload README.md with huggingface_hub

README.md

@@ -1349,4 +1349,49 @@ configs:
   data_files:
   - split: test
     path: full/test-*
+license: cc-by-4.0
+tags:
+- concept-bottleneck-models
+- phishing-detection
+- user-study
 ---
+# NWeak/CBM-user-study-emails
+
+Participant responses from the emails concept-bottleneck user study. Three configs:
+
+- **`full`** (417 participants, one row each): every participant, as
+  originally recorded. `ExperimentalCondition`: `LabelOnly`,
+  `NonInteractiveConcepts`, `InteractiveConcepts`, or `NoSupport`.
+- **`clean_wide`** (363 participants, one row each): `full` filtered to
+  remove participants who were (a) blank/incomplete, (b) failed
+  `AttentionCheck1`/`AttentionCheck2`, or (c) switched away from the
+  browser tab more than 3 times total across the 10 stimuli (a
+  tab-switching risk heuristic, e.g. possible use of an external tool to
+  answer). Kept participants' values are otherwise byte-identical to
+  `full`, except stimulus 268's columns are blanked in place (slot position
+  preserved) for participants who saw it -- see the 263/268 note below.
+- **`clean_long`** (3539 rows, one row per participant-stimulus, main 10
+  stimuli only): the same 363 participants as `clean_wide`, melted to long
+  format. `TimeSpent` is in minutes (`raw_ms / 60000`);
+  `Confidence_Adjusted = abs(Confidence - 7)` folds the 1-13 confidence
+  scale into a direction-independent 0-6 measure. Stimulus 268 rows are
+  fully absent here (not just blanked).
+
+**Known issue (stimuli 263/268)**: due to a human error, the model
+activations actually shown to participants for stimuli 263 and 268 did not
+match the model's real output. 268 also flipped which concepts were active
+(materially different information shown) -- all `full` observations for
+268 are blanked out in `clean_wide`/`clean_long`. 263's error preserved the
+same binary concept pattern (values differ, signs don't), so its
+observations were kept; its GT is correct, only its recorded activations
+are unreliable. Full root-cause writeup in this repo's `emails/DATASET_NOTES.md`
+and `emails/PARTICIPANT_DATA_REPORT.md`.
+
+This is our own contribution (participant responses, not the underlying
+email stimuli), so it's licensed independently -- see
+[`NWeak/emails-mirror`](https://huggingface.co/datasets/NWeak/emails-mirror)
+for the email data itself and
+[`NWeak/emails-user-study`](https://huggingface.co/datasets/NWeak/emails-user-study)
+for our model's own computed activations/predictions (a different dataset
+from this one -- that one is model output, this one is participant
+responses).