hunterbown/bell-labs-technical-archive

Bell Labs Documents and Stuff

This is a conservative public-release subset of the internal BELLA continued-pretraining corpus. It keeps the Bell-system technical material that survived a stricter final pass for public dataset hosting and removes records that still looked risky, off-scope, or too low-signal for a Hugging Face corpus listing.

What is in the release

Split	Documents
train	1220
validation	29
test	42

The release contains 1291 documents out of 1530 internally curated pretraining documents.

Document types:

• journal_issue: 875
• technical_report: 370
• patent: 23
• book: 12
• manual: 11

Source families:

• archive_org: 1265 documents
• google_patents: 23 documents
• other: 3 documents

Selected extraction backends among kept documents, when recoverable from the current local catalog:

• unknown: 1242
• pdftotext: 48
• archive_text: 1

Character counts after cleanup:

• min: 2514
• median: 114491
• max: 2734561

Where the data came from

Most documents were pulled from public Archive.org item pages that host Bell System Technical Journal issues, Bell Laboratories Record issues, Bell System Practices, Bell System / Western Electric technical manuals, and Bell-system-adjacent engineering reports. Patent records in this subset were pulled from public Google Patents pages.

Every released row includes per-document provenance fields such as source_url, archive_ref, and source_family so downstream users can trace each text file back to the public item page that it came from. A selected_extraction_backend field is also present, but many older rows remain unknown because parts of the internal corpus were built before backend tracking was recorded uniformly in SQLite.

How the corpus was built

1. Source records were discovered and imported into the local bella.db catalog.
2. For Archive.org-backed items, the pipeline downloaded the preferred PDF and any usable Archive.org text derivative.
3. Text extraction ran quality-first rather than single-backend-first:
   • pdftotext first when a PDF existed
   • Archive.org DjVu text as an alternate derivative when available
   • optional Qianfan OCR fallback only when pdftotext looked weak or mixed
4. Page-level heuristics removed obvious junk such as library stamps, scan boilerplate, table-of-contents pages, index pages, references, HTTP/header dumps, OCR markup artifacts, and pages with too little body text.
5. Document-level cleanup trimmed leading frontmatter such as Google/JSTOR boilerplate, issue mastheads, and leading OCR noise.
6. This public release applied one more conservative pass to exclude records with explicit restriction language, trade-secret notices, off-scope government/legal material, table-of-contents or index-only records, and very short bodies.

Final public-release exclusions

The final pass removed 239 documents. Exclusion counts by reason:

• index_or_ordering_title: 69
• short_body_under_min_chars: 55
• government_archive_title: 39
• restricted_reproduction_notice: 35
• personal_noncommercial_notice: 22
• trade_secret_notice: 20
• post_1990_non_patent: 10
• table_of_contents_title: 9
• consumer_magazine_title: 8
• all_rights_reserved_notice: 5
• offscope_misc_title: 3
• oral_history_title: 2
• legal_case_title: 1

The full exclusion log is in meta/excluded_records.jsonl.

OCR and extraction notes

This is not a hand-transcribed corpus. It is a cleaned OCR/text-extraction corpus. Some documents were born-digital or extracted cleanly with pdftotext; others depend on OCR or Archive.org text derivatives. The text is usable for corpus work, but it is not guaranteed to be page-faithful, typo-free, or complete.

Important limitations:

• OCR noise still exists in places, especially in older scans and diagram-heavy technical material.
• The corpus is Bell-focused, not a complete Bell Labs bibliography.
• The release process is conservative, but it is not legal advice.
• Some metadata fields were inferred or normalized during ingestion and cleanup.

Intended use

This subset is appropriate for:

• continued pretraining or domain adaptation experiments
• retrieval, search, and corpus analysis over Bell-system technical writing
• historical telecom and computing research where OCR noise is acceptable

This subset is not appropriate for:

• licensing-sensitive redistribution without your own review of the source items
• claims of perfect OCR fidelity
• high-stakes factual applications without source verification

Files

• data/train.jsonl, data/validation.jsonl, data/test.jsonl: the Hugging Face-ready data splits
• meta/release_manifest.json: build summary, counts, and checksums
• meta/excluded_records.jsonl: records removed by the public-release filter
• CHECKSUMS.sha256: file hashes for the whole release directory

Method provenance

This package was generated from:

• internal curated input: data/release/bella_v1/pretrain.jsonl
• build script: scripts/build_hf_corpus_release.py
• local SQLite catalog: data/bella.db
• release directory: data/release/bell_labs_documents_and_stuff

Citation

If you use the corpus, cite the dataset repo plus the original source repositories named in each row's provenance fields.