# PRAGMA Data

The released PRAGMA data is hosted on [Hugging Face](https://huggingface.co/datasets/stellahj/PRAGMA). The large JSON artifacts are not stored in the GitHub repository.

From the repository root, download the released files into this directory with:

```bash
python -m pip install -U huggingface_hub
hf download stellahj/PRAGMA \
  --repo-type dataset \
  --include "data/*.json" \
  --local-dir .
```

The download provides:

- `full_sessions.json`: complete timestamp-ordered conversation histories.
- `metadata.json`: benchmark queries and gold evidence-session indices.
- `response_metrics.json`: released alignment and grounding rubrics.
- `evid_sessions.json`: event and trajectory evidence conversations.
- `filler_sessions.json`: unrelated filler conversations.
- `query_data.json`: intermediate benchmark specifications.
- `privasis.json`: sampled source-user attributes used during construction.

The construction pipeline also writes regenerated artifacts to this directory.