README.md

metadata

pretty_name: SEC 10-K Pre-processed
language:
  - en
license: other
task_categories:
  - translation
tags:
  - finance
  - sec
  - translation

SEC 10-K Pre-processed

This dataset contains pre-processed SEC 10-K English text with sentence/chunk/paragraph files for translation-oriented use.

Source

• Original dataset: PleIAs/SEC
• This release is a pre-processed derivative of that source (cleaning, filtering, and deduplication).

Splits / Files

• sentence_track.jsonl: 1,620,583 rows
• chunk_track.jsonl: 914,217 rows
• paragraph_track.jsonl: 5,367,690 rows
• translation_dataset.jsonl: 7,902,490 rows (concatenated)

Columns

Shared columns: document_id, filing_id, filename, report_year, fiscal_year, cik, company_name, fiscal_year_end, form_type, section, section_title, section_id, paragraph_id, paragraph_index, source_text, word_count, is_boundary, boundary_reason, candidate_id, granularity, source_paragraph_id.

Track-specific columns:

• sentence: sentence_index_in_paragraph
• chunk: sentence_start_in_paragraph, sentence_end_in_paragraph, chunk_sentence_count
• paragraph: no extra columns

Recommended Loading

from datasets import load_dataset

ds = load_dataset(
    "json",
    data_files={
        "sentence": "sentence_track.jsonl",
        "chunk": "chunk_track.jsonl",
        "paragraph": "paragraph_track.jsonl",
    },
)
print(ds)

Deduplication

Deduplication was applied in two stages.

1. Paragraph master stage (upstream):

• Exact dedup by normalized text hash: lowercase + whitespace-normalized text.
• Approximate dedup with MinHash LSH (threshold=0.7, num_perm=128, 5-token shingles).

2. Translation track stage (this release):

• Track-level exact dedup by normalized text hash within each track (sentence, chunk, paragraph).
• If a paragraph candidate is selected, chunk/sentence candidates from that source paragraph are skipped.
• If a chunk candidate is selected, sentence candidates covered by that chunk span are skipped.

Notes

• The source is SEC 10-K text; users are responsible for downstream compliance checks.
• translation_dataset.jsonl duplicates rows already present in the three track files. For Hub release, uploading only track files is usually enough.