| --- |
| license: cc-by-nc-4.0 |
| language: |
| - en |
| pretty_name: MatrixNorm BVA AMA-Era 170K |
| size_categories: |
| - 100K<n<1M |
| annotations_creators: |
| - machine-generated |
| language_creators: |
| - found |
| multilinguality: |
| - monolingual |
| source_datasets: |
| - original |
| tags: |
| - legal |
| - veterans |
| - bva |
| - board-of-veterans-appeals |
| - administrative-law |
| - legal-ai |
| - public-records |
| - evidence-extraction |
| - issue-level-outcomes |
| - citations |
| - matrixnorm |
| task_categories: |
| - text-classification |
| - token-classification |
| - feature-extraction |
| - text-retrieval |
| --- |
| |
| # MatrixNorm BVA AMA-Era 170K |
|
|
| **MatrixNorm BVA AMA-Era 170K** is a normalized legal-AI dataset built from public |
| Board of Veterans' Appeals (BVA) decision records from the AMA era. |
|
|
| This dataset converts messy public veterans-law decisions into structured, |
| AI-ready legal records with extracted issues, outcomes, conditions, citations, |
| judges, appeal-lane signals, metadata, and reasoning fields. |
|
|
| **Current release: ~181K rows — but ≈108K *unique* decisions** (this file contains |
| duplicate rows; see the data-quality note below). A cleaned, de-duplicated version |
| is in progress. |
|
|
| This dataset is part of the broader **MatrixNorm** project: turning messy public |
| records into structured, AI-ready evidence graphs. |
|
|
| --- |
|
|
| ## ⚠️ Data-quality note — duplicate rows (cleaned version coming) |
|
|
| **This file currently contains duplicate rows.** It has **181,024 rows but only |
| 107,817 unique decisions** — roughly **73,000 rows are exact, byte-for-byte |
| duplicates**. The file was concatenated from per-year source CSVs that **overlapped |
| on 2021 and 2022**, so those years are repeated ~2–3×. De-duplicating on the |
| decision id collapses the file to its true size with **no loss** (the extra rows |
| are identical copies). |
|
|
| **Unique decisions by year (after de-duplication):** |
|
|
| | Year | Unique decisions | |
| |---|---| |
| | 2019 | 3,905 | |
| | 2020 | 19,476 | |
| | 2021 | 20,638 | |
| | 2022 | 26,287 | |
| | 2023 | 37,512 | |
| | **Total (2019–2023)** | **107,817** | |
|
|
| **If you train on this file as-is:** de-duplicate on the decision id first, and use |
| a **grouped split by decision id** so duplicate (and multi-issue) rows don't leak |
| across train/test. |
|
|
| **📌 An updated version is coming.** A **cleaned, de-duplicated** release will |
| supersede this file. Watch this repo for the update. |
|
|
| --- |
|
|
| ## Overview |
|
|
| Board of Veterans' Appeals decisions contain valuable legal and administrative |
| reasoning, but they are difficult to use directly because the source records are |
| long, inconsistent, and mostly unstructured. |
|
|
| MatrixNorm transforms those public records into structured legal data designed for |
| research, retrieval, analytics, and legal-AI workflows. |
|
|
| The goal is to make AMA-era BVA decisions easier to use for: |
|
|
| - legal AI research |
| - veterans-law analytics |
| - issue-level outcome modeling |
| - citation extraction |
| - grant / deny / remand classification |
| - condition extraction |
| - appeal-lane analysis |
| - judge-level pattern analysis |
| - reasoning-pattern analysis |
| - public-record normalization benchmarks |
|
|
| --- |
|
|
| ## Why this matters |
|
|
| Most legal datasets are either raw text collections or broad case-law corpora. |
|
|
| This dataset focuses on a specific, high-impact administrative-law domain: |
| **veterans benefits appeals.** |
|
|
| Instead of only providing full decision text, MatrixNorm extracts structured |
| fields from BVA decisions, making the data more useful for search, analysis, model |
| training, evaluation, and research. |
|
|
| --- |
|
|
| ## Dataset status |
|
|
| This release contains approximately: |
|
|
| - **170K** normalized AMA-era BVA records |
| - structured issue and outcome fields |
| - claimed-condition extraction |
| - judge and decision metadata |
| - citation extraction |
| - appeal-lane and procedural signals |
| - reasoning-pattern fields |
|
|
| A cleaned, de-duplicated version of this dataset is in progress (see the |
| data-quality note above). |
|
|
| --- |
|
|
| ## Fields |
|
|
| Exact columns may vary by release file, but fields may include: |
|
|
| | Field | Description | |
| |---|---| |
| | `decision_id` | Unique decision or source identifier | |
| | `decision_date` | Date of the BVA decision when available | |
| | `docket_number` | Docket or appeal identifier when available | |
| | `appeal_lane` | AMA lane or procedural signal when extracted | |
| | `judge` | Veterans Law Judge name when extracted | |
| | `issue_text` | Raw or cleaned issue statement | |
| | `condition` | Normalized claimed condition or disability | |
| | `outcome` | Issue-level outcome label | |
| | `outcome_group` | Grant / deny / remand / dismiss style grouping | |
| | `citations` | Extracted legal citations | |
| | `statutes` | Extracted statutory references | |
| | `regulations` | Extracted regulatory references | |
| | `evidence_terms` | Extracted evidence or medical/legal signal terms | |
| | `reasoning_pattern` | Normalized reasoning or decision pattern | |
| | `source_text` | Source decision text or excerpt, depending on release | |
| | `metadata` | Additional extracted decision metadata | |
|
|
| --- |
|
|
| ## Suggested use cases |
|
|
| ### Legal AI research |
|
|
| Use the dataset to study administrative-law decision patterns in veterans benefits |
| appeals. |
|
|
| Possible tasks: |
|
|
| - issue classification |
| - outcome classification |
| - condition extraction |
| - citation extraction |
| - procedural-lane analysis |
| - decision-pattern clustering |
| - legal retrieval |
| - metadata enrichment |
|
|
| ### Veterans-law analytics |
|
|
| Potential research questions: |
|
|
| - Which claimed conditions appear most often? |
| - Which issues are most frequently remanded? |
| - Which citations appear most often in grants, denials, and remands? |
| - How do AMA appeal-lane signals relate to outcomes? |
| - Which reasoning patterns repeat across decisions? |
| - What conditions produce the highest remand frequency? |
|
|
| ### Legal data normalization |
|
|
| This dataset can also be used to test extraction and normalization pipelines for |
| messy public legal records. |
|
|
| Possible evaluation tasks: |
|
|
| - issue segmentation |
| - citation parsing |
| - condition normalization |
| - judge extraction |
| - date extraction |
| - procedural metadata extraction |
| - outcome labeling |
|
|
| --- |
|
|
| ## Data source |
|
|
| The underlying decisions come from public Board of Veterans' Appeals records. |
|
|
| MatrixNorm adds value by normalizing, structuring, extracting, labeling, and |
| organizing the public decision content into AI-ready records. |
|
|
| Raw BVA decision text originates from public U.S. government records. |
| MatrixNorm's normalized schema, extracted fields, annotations, metadata, issue |
| labels, citation parsing, reasoning-pattern fields, and dataset organization are |
| released under the license listed below. |
|
|
| --- |
|
|
| ## License |
|
|
| This dataset is released under **CC BY-NC 4.0**. |
|
|
| You may use it for research, education, nonprofit, and other non-commercial |
| purposes **with attribution**. |
|
|
| **Commercial use is not permitted under this public license.** |
| For commercial licensing, partnerships, or enterprise access, contact the dataset |
| maintainer. |
|
|
| **Suggested attribution:** |
|
|
| > MatrixNorm BVA AMA-Era 170K by William T. Miller / MatrixNorm. |
| > Normalized public Board of Veterans' Appeals records for legal-AI research. |
| > Released under CC BY-NC 4.0. |
|
|
| --- |
|
|
| ## Limitations |
|
|
| This dataset is automatically normalized and may contain extraction errors. |
|
|
| Known limitations may include: |
|
|
| - imperfect issue segmentation |
| - imperfect condition normalization |
| - missing or inconsistent judge extraction |
| - citation parsing errors |
| - mixed outcomes inside a single decision |
| - legacy appeal records appearing inside AMA-era date ranges |
| - incomplete appeal-lane detection |
| - OCR or source-formatting artifacts |
| - labels that may require human review for high-stakes use |
| - **duplicate rows in the current file** (≈73K exact-duplicate rows; ~108K unique |
| decisions) — see the data-quality note near the top; a cleaned version is coming |
|
|
| **Do not treat this dataset as legal advice.** Do not use model outputs based on |
| this dataset as a substitute for an attorney, accredited representative, Veterans |
| Service Organization, or official VA decision review. |
|
|
| --- |
|
|
| ## Project direction |
|
|
| MatrixNorm is building a network of normalized public-record datasets across |
| legal, health, safety, food, environmental, and government domains. |
|
|
| The goal is to create structured evidence data for: |
|
|
| - AI training |
| - AI evaluation |
| - retrieval systems |
| - risk analysis |
| - public-interest research |
| - domain-specific legal and administrative analytics |
|
|
| MatrixNorm turns messy public records into usable AI infrastructure. |
|
|
| --- |
|
|
| ## Maintainer |
|
|
| **William T.L Miller — MatrixNorm** |
|
|
| Hugging Face: [williamTLmiller](https://huggingface.co/williamTLmiller) |
|
|
| Commercial licensing, collaborations, and research partnerships are available by |
| request. |
|
|
