CoolFace
Datasetpublic

andrew-mitchel/private-letter-rulings

Private Letter Rulings Text of IRS Private Letter Rulings (and other written determinations [TAMs, CCAs, etc.]), covering 1999 through August 2026. The IRS publishes these as PDF files each week; these were converted to text using pdfminer, falling back to OCR via pytesseract where needed. Dataset Structure 45,401 rows, one per ruling. Columns: Column Type Description wd_number string 9-digit IRS written determination number: 4-digit year + 2-digit week… See the full description on the dataset page: https://huggingface.co/datasets/andrew-mitchel/private-letter-rulings.

sourceHugging Facemitupdated 24d agoView on Hugging Face
0likes92downloads
Dataset Card

Private Letter Rulings

Text of IRS Private Letter Rulings (and other written determinations [TAMs, CCAs, etc.]), covering 1999 through August 2026. The IRS publishes these as PDF files each week; these were converted to text using pdfminer, falling back to OCR via pytesseract where needed.

Dataset Structure

45,401 rows, one per ruling. Columns:

ColumnTypeDescription
wd_numberstring9-digit IRS written determination number: 4-digit year + 2-digit week of release + 3-digit sequence within that week (e.g. 199901001 = 1999, week 01, ruling 1)
textstringFull text extracted from the PDF

Dataset Statistics

  • —Rows: 45,401
  • —Year range: 1999–2026 (2026 partial, through week 32)
  • —Rulings per year: ranged roughly 1,850–2,800/year from 1999–2010, declined steadily to 800–950/year in recent years (2018–2025)
  • —Text length: median ~8,900 characters, mean ~12,300 characters, max ~337,000 characters
  • —Empty or non-substantive text: 84 rows have no text at all, and a further ~220 rows contain only a short placeholder (e.g. PULLED (Inspection Delay), PULLED (Duplicate)) instead of ruling text — see Known Limitations

Usage

python
from datasets import load_dataset

ds = load_dataset('andrew-mitchel/private-letter-rulings')

Or locally with pandas:

python
import pandas as pd

df = pd.read_parquet('plrs.parquet')

Source & Processing

Source PDFs are the IRS's weekly Written Determinations releases. Text was extracted from each PDF and lightly cleaned:

  • —All newlines and form feeds were removed (each row's text is a single line).
  • —Section marks (§) are preserved.

Known Limitations

  • —84 rows have empty text, and roughly 220 additional rows contain only a short IRS placeholder string (e.g. PULLED (Inspection Delay), PULLED (Duplicate)) rather than ruling text — these correspond to documents the IRS pulled from its weekly release after initial posting.
  • —Text was extracted directly from PDFs and may include OCR-era artifacts (stray characters, watermark/page-number fragments) inherited from the source documents, particularly in older filings.
  • —wd_number is the IRS's own written determination number, not a sequential row index; it is unique per row but not contiguous (some numbers in a given week were never issued or were later pulled).
  • —This dataset is for research and informational use only; it is not legal advice. Private letter rulings apply only to the taxpayer who requested them and may not be cited or relied upon as precedent (IRC § 6110(k)(3)).