Fix dataset card: use train split so load_dataset and the viewer work; fix schema link; add Croissant and citation sections
Browse files
README.md
CHANGED
|
@@ -12,11 +12,11 @@ size_categories:
|
|
| 12 |
configs:
|
| 13 |
- config_name: full_dataset
|
| 14 |
data_files:
|
| 15 |
-
- split:
|
| 16 |
path: full_dataset.csv
|
| 17 |
- config_name: sampled
|
| 18 |
data_files:
|
| 19 |
-
- split:
|
| 20 |
path: dataset.csv
|
| 21 |
---
|
| 22 |
|
|
@@ -39,6 +39,8 @@ deterministically from `(user_id, domain)` via UUID5.
|
|
| 39 |
|---|---|---|
|
| 40 |
| `full_dataset.csv` | 2,000 | All users from all five studies, no filtering or sampling. |
|
| 41 |
| `dataset.csv` | 1,000 | Sampled to 200 users per domain (healthcare filtered to records with informative previous-experience values), `random_state=42`. |
|
|
|
|
|
|
|
| 42 |
|
| 43 |
## Schema notes
|
| 44 |
|
|
@@ -57,21 +59,34 @@ prefs = ast.literal_eval(df.iloc[0]["preferences"])
|
|
| 57 |
```
|
| 58 |
|
| 59 |
A column-by-column schema for every file is documented at
|
| 60 |
-
[docs/SCHEMA.md](https://github.com/
|
| 61 |
in the accompanying code repository.
|
| 62 |
|
| 63 |
## Loading
|
| 64 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 65 |
```python
|
| 66 |
from datasets import load_dataset
|
| 67 |
|
| 68 |
# Full unfiltered dataset (2,000 users)
|
| 69 |
-
ds = load_dataset("jamesflemings/PrivacySIM", "full_dataset", split="
|
| 70 |
|
| 71 |
-
# 1,000-user
|
| 72 |
-
ds = load_dataset("jamesflemings/PrivacySIM", "sampled", split="
|
| 73 |
```
|
| 74 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 75 |
## Source studies
|
| 76 |
|
| 77 |
The data is derived from five public user studies. Each row's `domain`
|
|
@@ -112,4 +127,17 @@ Full attribution, license texts, copyright notices, and indication of
|
|
| 112 |
modifications for every source study are provided in the
|
| 113 |
[`THIRD_PARTY_LICENSES.md`](./THIRD_PARTY_LICENSES.md) file shipped alongside
|
| 114 |
the dataset. Anyone redistributing or building on this dataset must preserve
|
| 115 |
-
that attribution file (or an equivalent).
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 12 |
configs:
|
| 13 |
- config_name: full_dataset
|
| 14 |
data_files:
|
| 15 |
+
- split: train
|
| 16 |
path: full_dataset.csv
|
| 17 |
- config_name: sampled
|
| 18 |
data_files:
|
| 19 |
+
- split: train
|
| 20 |
path: dataset.csv
|
| 21 |
---
|
| 22 |
|
|
|
|
| 39 |
|---|---|---|
|
| 40 |
| `full_dataset.csv` | 2,000 | All users from all five studies, no filtering or sampling. |
|
| 41 |
| `dataset.csv` | 1,000 | Sampled to 200 users per domain (healthcare filtered to records with informative previous-experience values), `random_state=42`. |
|
| 42 |
+
| `croissant.json` | | Croissant 1.1 metadata with Responsible AI (RAI) fields describing both CSV files. |
|
| 43 |
+
| `THIRD_PARTY_LICENSES.md` | | Attribution, license texts, and modification notices for the five source studies. |
|
| 44 |
|
| 45 |
## Schema notes
|
| 46 |
|
|
|
|
| 59 |
```
|
| 60 |
|
| 61 |
A column-by-column schema for every file is documented at
|
| 62 |
+
[docs/SCHEMA.md](https://github.com/james-flemings/PrivacySIM/blob/main/docs/SCHEMA.md)
|
| 63 |
in the accompanying code repository.
|
| 64 |
|
| 65 |
## Loading
|
| 66 |
|
| 67 |
+
Each configuration exposes a single `train` split containing every row of the
|
| 68 |
+
corresponding CSV file (the data is evaluation-only; the split name is a
|
| 69 |
+
Hub convention, not a train/test partition).
|
| 70 |
+
|
| 71 |
```python
|
| 72 |
from datasets import load_dataset
|
| 73 |
|
| 74 |
# Full unfiltered dataset (2,000 users)
|
| 75 |
+
ds = load_dataset("jamesflemings/PrivacySIM", "full_dataset", split="train")
|
| 76 |
|
| 77 |
+
# 1,000-user sample used in the paper
|
| 78 |
+
ds = load_dataset("jamesflemings/PrivacySIM", "sampled", split="train")
|
| 79 |
```
|
| 80 |
|
| 81 |
+
## Croissant metadata
|
| 82 |
+
|
| 83 |
+
A [Croissant](https://mlcommons.org/croissant/) 1.1 JSON-LD file that
|
| 84 |
+
describes both CSV files, their fields, the license, and the Responsible AI
|
| 85 |
+
(RAI) and provenance metadata is shipped as
|
| 86 |
+
[`croissant.json`](./croissant.json). The Hub also serves auto-generated
|
| 87 |
+
Croissant metadata at
|
| 88 |
+
`https://huggingface.co/api/datasets/jamesflemings/PrivacySIM/croissant`.
|
| 89 |
+
|
| 90 |
## Source studies
|
| 91 |
|
| 92 |
The data is derived from five public user studies. Each row's `domain`
|
|
|
|
| 127 |
modifications for every source study are provided in the
|
| 128 |
[`THIRD_PARTY_LICENSES.md`](./THIRD_PARTY_LICENSES.md) file shipped alongside
|
| 129 |
the dataset. Anyone redistributing or building on this dataset must preserve
|
| 130 |
+
that attribution file (or an equivalent).
|
| 131 |
+
|
| 132 |
+
## Citation
|
| 133 |
+
|
| 134 |
+
If you use PrivacySIM, please cite the paper (and the source studies above):
|
| 135 |
+
|
| 136 |
+
```bibtex
|
| 137 |
+
@inproceedings{flemings2026privacysim,
|
| 138 |
+
title = {{PrivacySim}: Evaluating {LLM} Simulation of User Privacy Behavior},
|
| 139 |
+
author = {Flemings, James and Annavaram, Murali},
|
| 140 |
+
booktitle = {Advances in Neural Information Processing Systems (NeurIPS), Evaluations and Datasets Track},
|
| 141 |
+
year = {2026}
|
| 142 |
+
}
|
| 143 |
+
```
|