Datasets:
FilingsBench Controls v0.1 candidate slice
This issuer-disjoint slice contains 12 SEC-filed primary HTML documents: ten 10-Ks and two 10-Qs. It balances five effective and five ineffective annual ICFR conclusions, and includes two quarterly cases where the correct behavior is to extract disclosure-controls effectiveness while abstaining on unsupported annual ICFR and material-weakness labels.
The four fields are icfr_effectiveness, disclosure_controls_effectiveness,
material_weakness_disclosed, and remediation_status. See
docs/controls_annotation_guidelines.md for exact
definitions and adversarial rules.
These labels are candidate_pending_independent_review. They were read from filed SEC HTML but
have not been independently double-annotated or adjudicated, so this slice is not described as
human-gold. The raw predictions remain separate from the labels.
Rebuild and evaluate with:
finstruct controls-dataset build
finstruct controls-benchmark run data/filingsbench_controls_v0.1/records.jsonl