Data and task composition
| Source | Tasks | Task documents | Remaining learning documents |
|---|---|---|---|
| amazon_beauty | 13 | 130,000 | 330,885 |
| app_reviews | 13 | 65,000 | 1,660 |
| cfpb | 12 | 120,000 | 483,788 |
| nhtsa | 12 | 120,000 | 128,135 |
| Total | 50 | 435,000 | 944,468 |
There are 20 group differences, 15 temporal changes and 15 compound associations.
Each source's corpus is deterministically divided into disjoint research cohorts.
The 50 briefs in benchmark/research_briefs.json express distinct investigation
objectives, not predefined answers. Broad objectives can overlap conceptually.
Task corpora are gzip JSONL. Common fields are doc_id, source, text, title, timestamp, timestamp_kind, entity_id, entity_name, category and rating. Available nonconstant source metadata may include state, make, model_year and store. report_year is derived from timestamp. Only task.allowed_metadata_fields can be used for population filters and metadata group selection. Free text remains untrusted author reports; timestamp describes the released date kind, not necessarily incident time. Null metadata is preserved.
The optional learning pool has 278 Parquet shards containing doc_id, source, text and title, with no annotations. App Reviews has a small remaining learning pool; source-balanced training is not implied. Evaluation documents were selected from a previously public curated learning snapshot. They are NOT guaranteed unseen. Current task IDs and learning IDs are disjoint; the curated input's normalized and conservative-template deduplication policy is inherited. Shared entities, authors and sources can remain; document separation is not independence.
Rebuild from the exact curated input and normalized source metadata:
python scripts/rebuild_benchmark.py --pool /path/to/input/learning \
--processed /path/to/normalized --output /new/output/directory
The builder requires a new output directory, records hashes of input shards and research briefs, and never invents reference annotations. Normalized inputs must have the schema used in the script; this is not an upstream raw-download parser. Released files and their manifest are sufficient to run the benchmark without reconstruction. Data provenance, source eligibility and redistribution terms are described in SOURCES.md. Do not infer incidence rates for products, vehicles or the wider population from these sampled reports.