docs/data_format.md

# Data format

Raw JSONL files have exactly nine fields; one line is one candidate/scoring unit.

| Field | Meaning |
|---|---|
| id | Stable global integer candidate ID. Not a row offset. |
| original_id | Stable global original-question ID shared across versions and attacks. |
| source | Source dataset key; see sources.md. |
| scenario | Task category; image sets use categories within each source benchmark. |
| clean | Complete original logical request, including state and all questions. |
| adv | Complete attacked logical request. |
| gt | Ground truth, preserving native string, boolean or number type. |
| task | mode, question_ids and, for Noul, labels/positive_label where applicable. |
| attack | Attack method or injection template name. |

The schema is in raw_record.schema.json. dataset_info.json lists the actual values and
counts for sources, scenarios, modes and attacks. IDs are not renumbered in subsets.
The same original may have multiple attack candidates and scoring targets. An image
original is a question, not merely a shared image. Future versions reuse the established
ID registry. No per-record validation, selection metadata or model outputs are included.

## JSONL and Hugging Face views

`raw/<config>/data.jsonl` retains native nested JSON. Paths in clean.images/adv.images
are relative to that raw configuration directory. Copy the images folder with the JSONL.
The Parquet view serializes clean, adv, gt and task as lossless JSON **strings**, because
requests and labels have heterogeneous schemas and types. Use scripts/io_utils.py to
decode them. The other five fields keep their native scalar types.

In the Dataset Viewer, image configurations show `clean_image` and `adv_image` first,
followed by scenario, attack and source. Text configurations show `clean` and `adv` first.
Stable IDs appear last. Column order does not change the native record or scoring.

The three image configurations additionally expose clean_image and adv_image, typed as
Hugging Face Image columns with embedded original PNG bytes. These are transport/display
helpers, not extra annotations. There is one image per side in this release. Image data
can be used without downloading the raw folder. No image was resized or re-encoded.

Retain all questions as input, including questions without GT. Only task.question_ids
participate in this record's scoring. Multiple Noul question IDs can jointly define one
classification. A question absent from question_ids is not necessarily unlabelled: it
may be scored by another record. The optional historical model field is transport metadata;
select the intended evaluation model explicitly. Supply images through its image interface.

Six GUI inputs contain historical file:// page URLs as literal stimulus text. Those URLs
are not external dependencies to fetch, and were not silently edited for this export.

## Loading and integrity checks

Install requirements.txt before using the scripts. The public Hub repository can be
loaded without authentication. To load a downloaded repository locally, use its directory
instead of the Hub repository ID.

The raw and Parquet views contain the same examples; do not concatenate them.
Run these commands from the downloaded repository directory:

```bash
python scripts/load_examples.py
python scripts/verify.py
```

The first script demonstrates loading; the second verifies checksums, record round trips
and bundled images. Neither calls an evaluation model. See evaluation.md and
scripts/scoring.py for native scoring and ASR definitions.