USAGE.md

# Usage

Load `tasks` and `candidates` with `split="full"` as shown in the README. Decode the content columns with `json.loads` and join the two tables on `task_id`.

| Fields | Meaning |
|---|---|
| `task_id`, `candidate_id` | Stable IDs linking tasks and candidates |
| `benchmark`, `task_type`, `source_*` | Source and task metadata |
| `task_spec_json`, `candidate_view_json` | Task specification and candidate content as JSON strings |
| `label`, `candidate_class` | PASS/FAIL reference label and construction category |
| `content_access` | Embedded content or AppWorld local reconstruction |
| `*_sha256`, `known_issue` | Content verification hashes and recorded issue annotations |

AppWorld content is null in Parquet. Obtain the official data using the [AppWorld setup instructions](https://github.com/StonyBrookNLP/appworld#installation), then run from this repository:

```bash
python -m pip install -r requirements.txt
python scripts/materialize.py --kind generation --appworld-root /path/to/appworld_root --output /path/to/tasks
python scripts/materialize.py --kind evaluation --appworld-root /path/to/appworld_root --output /path/to/candidates
```

The AppWorld root must contain `data/`. Use separate output directories outside this repository. The scripts reconstruct all 222 tasks and 4,993 candidates and reject source data that does not match the frozen hashes. Generate verifiers without access to candidates or labels. Keep reconstructed AppWorld plaintext local; its protected content requires encryption for public redistribution.