williamTLmiller/Bva_ama_decisions__structured_2019

BVA AMA Decisions — Structured (A-prefix)

Structured annotations over 5,990 U.S. Board of Veterans' Appeals (BVA) decisions issued under the Appeals Modernization Act (AMA) — the modernized appeals system that replaced the legacy process. Every record is an A-prefix docket decision (A21……), so this is a 100% AMA corpus: the part of the BVA record that the standard academic/legacy datasets do not cover.

Why this dataset is different

The widely-used BVA legal-NLP datasets are built on the legacy appeals system. AMA decisions use a different docket scheme (A-prefix IDs) and a structure the legacy system never had — most importantly the review lane the veteran selected. This dataset captures that:

• —`ama_docket` — the AMA review lane, present on 77% of decisions:
• —direct_review — 3,061 (51%)
• —hearing — 864 (14%)
• —evidence_submission — 737 (12%)
• —(not stated in the decision text) — 1,328 (22%)

This lane choice exists only in the AMA system and is not available in legacy corpora.

What's in it

• —Decisions: 5,990 (one row each), all regime = ama, all A-prefix dockets
• —Outcomes (notably balanced — an AMA signal): remanded 2,246 · denied 2,217 · granted 2,171 · dismissed 779 · reopened 12 · withdrawn 7. Unlike the legacy corpus (heavily remand-dominated), the modernized lanes produce far more final grants/denials — a real structural difference, not a relabel.
• —Conditions: musculoskeletal (back, knee), sensory/neurological (hearing loss, tinnitus, peripheral neuropathy), and psychiatric (PTSD) lead, as in the broader BVA record.

Columns

doc_id, source_url, issues, conditions, conditions_raw, condition_other, conditions_detailed, outcomes, outcome_by_issue, reasoning_by_issue, reasoning_completeness, reasoning_unfillable, regime, ama_docket, citations, evidence, reasoning_atoms, judge_dates, char_length

Field coverage

issues 98% · outcomes 97% · outcome_by_issue 97% · citations 100% · evidence 91% · reasoning_atoms 90% · reasoning_by_issue 85% · conditions 74% (the remainder are procedural/non-medical decisions; conditions_raw carries rare/unmapped diagnoses as free text).

Provenance, license & accuracy

• —Source: va.gov BVA decisions — works of the U.S. government, in the public domain. Every row links to its source_url for independent verification; the A-prefix doc_id is the AMA docket identity.
• —Scope note: this release is the 2021-docketed (A21) AMA cohort. AMA decisions of other docket years exist and can be added; this set stands on its own as a complete A21 corpus.
• —License: annotations and structured compilation © the author, released under CC-BY-NC-4.0 (free for research/non-commercial use; contact for commercial licensing). The underlying decisions are public domain.
• —Accuracy: high-quality machine annotations (extraction benchmarked at ~96% on a silver/LLM-labeled set; human gold verification ongoing). Treat as training data, not certified ground truth. No guaranteed predictions — outcomes are historical extractions, not forecasts.

Intended use

Training and evaluating legal extraction and outcome models on the modernized (AMA) appeals system specifically, AMA-vs-legacy comparison, docket-lane analysis, and veterans-claims tooling built for the current appeals regime.