--- language: - en license: other tags: - economics - finance - education - news - cpt - qwen size_categories: - 100K` artifact split/cleanup - Bloomberg contact boilerplate tail removal - table/separator noise removal (`====`, `YOY%` 등) - leading/trailing sentence-fragment trimming 3. Exact dedup 4. Sentence-aware chunking - tokenizer: `Qwen/Qwen3.5-4B` - `max_tokens=512`, `overlap_tokens=32` - docs with any `hard_split` dropped ## Latest Build (2026-03-31) From `external_econ_clean_summary.json` and `external_econ_chunks_sentence_512_32_summary.json`: - docs_after_exact_dedup: **123,336** - docs_chunked: **123,164** - docs_split: **48,191** - chunks_total: **205,614** - avg_chunks_per_doc: **1.6694** - docs_dropped_hard_split: **172** - hard_split_chunks in output: **0** ## Files - `external_econ_dedup.jsonl` - `external_econ_clean_summary.json` - `external_econ_chunks_sentence_512_32.jsonl` - `external_econ_chunks_sentence_512_32_summary.json` ## Notes - 원본 소스별 라이선스/사용조건은 각각 상이할 수 있습니다. - 실제 학습/배포 전, 소스별 라이선스와 사용 정책을 반드시 확인하세요.