beyond369/us-regulations-16k-64k

Dataset Card for Long US Regulations (16k–64k)

Dataset Summary

Long US Regulations (16k–64k) is a curated, high-integrity long-context corpus derived from the primary text of the United States Code of Federal Regulations (CFR). It consists of 793 extensive regulatory documents totaling 41,227,048 tokens, with an average document length of 51,988 tokens (measured using OpenAI's cl100k_base tokenizer).

Every document strictly spans between 16,384 and 65,536 tokens. Instead of arbitrary windowing or fixed token chunking, documents are constructed along natural statutory boundaries—preserving complete regulatory units (intact chapters or contiguous statutory parts). The corpus has undergone weighted downsampling across the Code of Federal Regulations, balancing domain diversity while capping single-title dominance to 4.0%.

---

Dataset Structure

Features

Field	Type	Description
chunk_id	string	Unique identifier (e.g., title_12_ch_II_partpack_1 or title_11_ch_I_COMPLETE).
cfr_title	string	CFR Title number (1 through 50, excluding reserved Title 35).
title_name	string	Official name of the CFR Title (e.g., Banks and Banking, Labor).
chapter	string	Statutory chapter identifier within the title.
unit_type	string	Packing container: complete_chapter or complete_parts_pack.
token_count	int64	Token count evaluated with cl100k_base (bounded in [16384, 65536]).
gov_weight	float64	Priority sampling weight assigned based on public governance tier.
text	string	Primary Markdown-formatted statutory text, assembled in legal hierarchy.

Summary Statistics

• Total Documents: 793
• Total Tokens: 41,227,048
• Token Bounds: 16,384 to 65,536 tokens
• Average Length: 51,988 tokens
• Active CFR Titles Represented: 48 titles (Title 35 is reserved in the CFR; Title 3 yielded no contiguous statutory units within the 16k–64k token boundary)
• Maximum Title Share: 31 documents (3.9% of corpus)

---

Source Data & Construction Pipeline

Source Data

The underlying statutory sections were sourced from the raw regulatory records in the docketx/us-regulations collection, which indexes section-level text and metadata from the U.S. Government Publishing Office (GPO) Code of Federal Regulations. This data is in the public domain.

Production Pipeline

The corpus was generated using a three-stage pipeline designed to eliminate boundary truncation:

1. Complete-Unit Packing (Statutory Preserving):

   • Sections were sorted using natural statutory ordering (title_num $\rightarrow$ chapter $\rightarrow$ part_num $\rightarrow$ appendix_num $\rightarrow$ section_num) to guarantee correct numerical and Roman-numeral progression.
   • Case 1 (Complete Chapters): If an entire administrative chapter measured between 16,384 and 65,536 tokens, it was preserved in its entirety (unit_type: complete_chapter).
   • Case 2 (Contiguous Part Packs): For chapters exceeding 65,536 tokens, sections were grouped by statutory Part. Parts were greedily packed in consecutive order until approaching the upper limit without splitting any individual Part (unit_type: complete_parts_pack). Incomplete trailing fragments below 16,384 tokens were discarded.
   • This initial packing yielded 1,258 candidate documents (65,295,352 total candidate tokens).

2. Governance-Tier Weighting: Titles were categorized into four public administration tiers to guide sampling:

   • Tier 1 (Weight 3.0): Essential public frameworks (Internal Revenue, Labor, National Defense, Judicial Administration, Aliens & Nationality).
   • Tier 2 (Weight 2.0): High-visibility operational governance (Banking, Public Health, Transportation, Food & Drugs, Energy).
   • Tier 3 (Weight 1.0): Specialized administration (Commerce, Wildlife, Navigation, Federal Acquisition).
   • Tier 4 (Weight 0.5): Administrative plumbing (General Provisions, Accounts).

3. Stratified Downsampling & Category Capping:

   • An iterative water-filling algorithm distributed target quotas across all CFR titles based on tier weights and available document pools.
   • A hard ceiling of 4% (max 31 documents per title) was enforced to prevent voluminous areas (such as Title 7 Agriculture, Title 40 Protection of Environment, or Title 12 Banking) from overwhelming the collection.
   • The final downsampled subset achieved the exact target of 793 documents with balanced topic entropy.

---

Potential Value & Use Cases

Continued Pre-Training (CPT)

Federal regulations exhibit distinctive characteristics that challenge language models: dense cross-references, deep hierarchical nesting, formal legal precision, and extensive conditional logic. Incorporating these documents into pre-training helps models develop structured deductive reasoning and long-range dependency tracking.

Diverse Administrative Law Domain Adaptation

Because of the tier-weighted sampling and strict 4% per-category ceiling, this dataset avoids overfitting to a single agency (such as the EPA or IRS). It comprehensively spans 48 titles—ranging from maritime shipping safety and patent rules to telecommunications and federal elections.

High-Quality Long-Context Evaluation & Needle Retrieval

Most synthetic long-context benchmarks stitch unrelated Wikipedia articles or fictional stories together. This corpus provides naturally occurring, hyper-dense, complex reference documents spanning 16k to 64k tokens. It is an ideal testbed for:

• Long-context retrieval-augmented generation (RAG).
• Multi-hop legal reasoning across interconnected regulatory subparts.
• Needle-in-a-Haystack (NIAH) evaluations on dense statutory definitions and compliance requirements.

---

Licensing & Ethics

• Text Rights: Code of Federal Regulations materials produced by the United States Federal Government are works of the public domain (U.S. Code Title 17, Section 105).
• Dataset License: Creative Commons Zero v1.0 Universal (CC0-1.0).