---
license: other
language:
- en
tags:
- preprocessing
- text-cleaning
- streaming
task_categories:
- text-generation
pretty_name: Bloomberg_Financial_News_processed
---

# Bloomberg_Financial_News_processed

## Source
- Original dataset: `danidanou/Bloomberg_Financial_News`
- Original dataset page: https://huggingface.co/datasets/danidanou/Bloomberg_Financial_News
- Split: `train`
- Config/subset: `(default)`

## Processing
- Streamed row-by-row with Hugging Face `datasets` (`streaming=True`).
- Text field extracted by priority from `--text-fields` (case-insensitive).
- Newline normalization: `\r\n` / `\r` -> `\n`.
- Removed non-visible control/format characters (`Unicode Cc/Cf/Cs/Co/Cn`, except `\n\t\r`).
- Converted non-breaking spaces (`U+00A0`) to regular spaces.
- Applied lightweight mojibake recovery heuristic for common UTF-8/cp1252 artifacts.
- Whitespace normalized (`[ \t]+` -> single space, 3+ newlines -> double newline).
- Dropped rows with cleaned text length `< 50`.
- Output schema: `text`, `language`, `style_tag`, `metadata`.
- Output shard format: `jsonl.gz`.

## Run Summary
- Last updated: 2026-04-25 06:45:12 UTC
- Seen rows: 446,762
- Kept rows: 446,707
- Dropped (no text): 0
- Dropped (too short): 55
- Changed rows: 433,182
- Length stats (chars): min=50, mean=2708.876, max=73941

## Notes
- `metadata` preserves source row context/fields for traceability.
- Processing is resumable with checkpoint files.