Update README.md

README.md

@@ -1,3 +1,123 @@
 ---
+---
 license: cc-by-nc-4.0
+language:
+- en
+pretty_name: BVA Structured Decisions (2019–2025)
+task_categories:
+- text-classification
+- token-classification
+- text-retrieval
+- question-answering
+tags:
+- legal
+- law
+- legal-nlp
+- veterans-affairs
+- disability
+- bva
+- structured-extraction
+- administrative-law
+size_categories:
+- 1K<n<10K
+---
+
+# BVA Structured Decisions (2019–2025)
+
+**Structured, issue-level records extracted from U.S. Board of Veterans' Appeals (BVA) decisions** — each decision parsed into its issues, conditions, outcomes, citations, and reasoning, with per-document provenance and completeness flags. Built for training and evaluating legal-AI models on veterans' disability adjudication.
+
+This is a **2,905-decision sample**, balanced across **seven years (2019–2025, ~415 decisions/year)**, so it's representative of the full corpus rather than skewed to one year. A larger full-corpus release and a commercial license are available (see **Access & licensing** below).
+
+> **Honesty note (please read).** Labels are **silver** (engine-produced, benchmarked against an LLM-labeled reference at ~96% outcome accuracy), **not** human-certified gold. Provenance and completeness ship with every row so you can verify and filter. See **Quality & accuracy**.
+
+---
+
+## What's in it
+
+| | |
+|---|---|
+| **Decisions** | 2,905 |
+| **Years** | 2019–2025 (~415 per year, balanced) |
+| **Format** | CSV (one row per decision) — 23 columns |
+| **Coverage** | 99% of rows carry extracted issues and citations; 96% carry reasoning atoms |
+| **Source** | Public BVA decisions published on va.gov |
+
+### Outcome distribution (decision-level)
+`remanded` 1,539 · `denied` 1,103 · `granted` 973 · `dismissed` 253 · `reopened` 68 · `withdrawn` 3
+
+### Reasoning completeness
+`full` 2,156 (74%) · `partial` 491 (17%) · `none` 258 (9%) — of the `none` rows, 108 are flagged `reasoning_unfillable` (the source letter genuinely contains no reasoning, so emptiness is correct, not a miss).
+
+### Appeals regime
+`legacy` 1,083 · `AMA` 60 · `unknown` 1,762 (tagged from the decision text, not the year).
+
+### Conditions
+**107 distinct medical conditions.** Top: back disability (490), knee disability (438), hearing loss (341), arthritis (302), psychiatric disorder (289), PTSD (287), peripheral neuropathy (272), sleep apnea (212), hypertension (195), foot disability (177).
+
+---
+
+## Column dictionary
+
+| Column | Description |
+|---|---|
+| `doc_id` | BVA citation/docket number (the 2-digit prefix is the fiscal year). |
+| `source_url` | Link to the original decision on va.gov for independent verification. |
+| `schema` | Extraction schema used (`bva`). |
+| `issues` | The appealed issues (pipe-separated). |
+| `conditions` / `conditions_raw` / `conditions_detailed` / `condition_other` | Canonical condition tokens, raw phrasing, laterality/qualifiers, and an out-of-vocabulary flag. |
+| `outcomes` | Decision-level dispositions (granted / denied / remanded / dismissed / reopened / withdrawn). |
+| `outcome_by_issue` | **Each issue tied to its own outcome** (the core training unit). |
+| `reasoning_by_issue` | Per-issue reasoning atoms (the "why"). |
+| `reasoning_completeness` | `full` / `partial` / `none` — quality tier for self-selecting a clean slice. |
+| `reasoning_unfillable` | `True` when the source letter has no reasoning at all (so `none` is expected). |
+| `regime` / `ama_docket` | Legacy vs. Appeals-Modernization-Act regime + AMA docket flag. |
+| `citations` | Statutory/regulatory citations (38 U.S.C. / 38 C.F.R.), normalized and deduped. |
+| `evidence` | Evidence types referenced (VA exam, private opinion, lay statement, etc.). |
+| `reasoning_atoms` | Canonical reasoning findings (nexus established/not, benefit-of-doubt, duty-to-assist, etc.). |
+| `judge_dates` | Decision date + Veterans Law Judge. |
+| `signals_extracted` / `matrix_cells_used` / `avg_route_score` / `char_length` | Extraction telemetry + raw length. |
+
+---
+
+## Why it's useful
+
+- **Issue-level supervision.** `outcome_by_issue` and `reasoning_by_issue` link each appealed issue to its disposition and rationale — not just a document-level label.
+- **Trainable + filterable.** Completeness tiers let you train on the clean `full` slice or use everything; `regime` lets you separate legacy vs. AMA.
+- **Verifiable.** Every row carries a `source_url` back to the public original.
+- **Representative.** Balanced across 2019–2025, so temporal/longitudinal splits are honest.
+
+**Intended uses:** training/evaluating models for claim-outcome prediction, issue and citation extraction, legal RAG over veterans' law, and fine-tuning assistants for BVA practice.
+
+---
+
+## Quality & accuracy (read before you rely on it)
+
+- **Silver, not gold.** Records are produced by a deterministic extraction engine and benchmarked against an LLM-labeled reference at **~96% outcome-extraction accuracy** (measured on 2019 issues). **Human-verified gold validation is in progress** and not yet reflected here.
+- **What's well-covered:** issues, conditions, outcomes, citations, and reasoning atoms (96–99% of rows populated).
+- **Known limits:** the `none`/`partial` reasoning rows are recoverable gaps, not curated blanks; `regime` is `unknown` for many rows where the text doesn't clearly signal it; condition extraction follows a controlled vocabulary (107 tokens) with an `condition_other` flag for the long tail.
+
+Treat the accuracy figure as a **silver benchmark**, and verify against `source_url` for any high-stakes use.
+
+---
+
+## Provenance, PII & ethics
+
+- **Source:** BVA decisions are U.S. federal records, published publicly on va.gov. The underlying text is public domain (17 U.S.C. §105); the **structured layer** (this dataset's value-add) is what the license below covers.
+- **PII:** BVA publishes decisions with appellants de-identified (referred to as "the Veteran"/initials). This release was screened with an automated PII gate (blocks SSN/phone/email/DOB/address). It may still contain incidental names (e.g., judges, place names) inherent to public legal text — review before any redistribution.
+- **Not legal advice.** This is research/ML data about adjudication patterns, not guidance for any individual claim.
+
+---
+
+## Access & licensing
+
+- **License: CC BY-NC 4.0** — free to use for **research and non-commercial** purposes, with attribution. The raw decision text is public domain; the **NC term applies to the structured annotations** in this dataset.
+- **Commercial use / full corpus:** the full multi-year corpus (and a commercial license) are available separately. **Contact the maintainer** to license it for commercial use.
+
+### Attribution
+> BVA Structured Decisions (2019–2025). Structured extraction of public Board of Veterans' Appeals decisions. Derived from va.gov public records; structured layer © the maintainer, released under CC BY-NC 4.0.
+
+---
+
+## Changelog
+- **v1** — 2,905 decisions, balanced 2019–2025; 23-column document schema; silver labels (~96% benchmark).
 ---