Situations paired with the assistance programmes that match them, the ones that do NOT match, and a matcher_notes field explaining the reasoning against published federal criteria. Every row carries never_guarantee_framing: the programme criteria 'appear relevant' rather than 'you qualify'. That constraint is the interesting part — it is a corpus built specifically so a model cannot promise a benefit it is not entitled to promise.

Provenance

This group merges 4 source corpora. Every row carries a _source_dataset field naming the file it came from, so the grouping is reversible:

aid_scenarios — 1,609 rows
aid_programs — 17 rows
aid_sources_ledger — 16 rows
cross_lingual_aid_corpus — 12 rows

Measured

Property	Value
Rows	1,654
Bytes	1,844,592
Source corpora merged	4
Distinct top-level fields	6

Counted with wc -l against the file staged here. Absolute machine paths were rewritten to portable tokens (<<TREES>>, <<WME>>, <<DRIVE>>, <<HOME>>) and the row count verified unchanged by that rewrite.

Known limitations

Synthetic and self-graded. Grades, judgements and notes come from the author's own pipeline, not independent annotation.
Small per source. Several constituent corpora are single-digit rows.
_source_dataset is added by this packaging, not present in the originals.
Not peer-reviewed or independently evaluated.
Security group only: the offensive rows are public MITRE-mapped techniques; the corpus's contribution is the pairing and the self-grading, not new offensive capability.

Licence

CatQualia Structural Isomorphism License (CSIL) v3.0 — attribution: Christopher Betances, catqualia.com. Non-commercial use is free with attribution; commercial use requires a commercial licence. Full text: LICENSE in this repository and catqualia.com/licensing

Downloads last month: 20

Total file size:

1.76 MB