| | --- |
| | license: mit |
| | --- |
| | PLOS Cleaned JSONL Dataset — 50+ GB of Machine-Ready Scientific Text |
| |
|
| | This repository delivers a fully cleaned and standardized large-scale corpus extracted from the complete PLOS scientific archive, reorganized and formatted into high-quality JSONL files. All content is optimized for immediate use in modern LLM training pipelines (OpenAI, Mistral, DeepSeek, Llama, Gemma, etc.). |
| | Total size (compressed): ~50 GB • Articles processed: hundreds of thousands • Cleaning quality: ~98–99% • Format: JSON Lines (.jsonl) |
| |
|
| | Features |
| |
|
| | Full PLOS corpus reorganized by year: |
| | output/datasets_by_year/year_2003.jsonl → year_2025.jsonl |
| |
|
| | Train/val/test splits ready for training: |
| | output/splits/train.jsonl, val.jsonl, test.jsonl |
| |
|
| | Grouped by journal (all PLOS branches): |
| | Medicine, Genetics, Pathogens, Climate, Computational Biology, Digital Health, Global Public Health, ONE, Biology, Neglected Tropical Diseases, etc. |
| | Example: |
| | output/datasets_by_journal/PLOS_Medicine.jsonl, PLOS_Genetics.jsonl, PLoS_Biology.jsonl, PLOS_ONE.jsonl, etc. |
| |
|
| | JSONL Structure |
| |
|
| | Each line follows the same cleaned schema, ideal for training, RAG, or vector embedding: |
| |
|
| | { |
| | "title": "Article title", |
| | "abstract": "Cleaned abstract...", |
| | "body_text": "Full cleaned article body...", |
| | "journal": "PLOS Medicine", |
| | "year": 2018, |
| | "authors": ["First Author", "Second Author"], |
| | "doi": "10.1371/journal.pmed.xxx", |
| | "keywords": ["biology", "health"], |
| | "clean_text": "Final merged cleaned text..." |
| | } |
| |
|
| | Cleaning Pipeline |
| |
|
| | Extraction from raw PLOS dumps |
| |
|
| | Multistage cleaning (regex, normalization, structure repair) |
| |
|
| | Removal of XML/HTML debris, broken text, figure callouts, corrupted sections |
| |
|
| | Unified schema + metadata alignment |
| |
|
| | Sorting by journal & by year |
| |
|
| | Automatic train/val/test splitting |
| |
|
| | Final QC ensuring no empty or malformed samples |
| |
|
| | Why This Dataset Matters |
| |
|
| | Massive 50+ GB scientific corpus ready for training |
| |
|
| | High-signal, peer-reviewed biomedical & scientific text |
| |
|
| | Perfect for: |
| |
|
| | biomedical/scientific LLM pretraining |
| |
|
| | SFT and supervised tasks |
| |
|
| | RAG systems & semantic search |
| |
|
| | reasoning models |
| |
|
| | dataset benchmarking |
| |
|
| | Fully stable, normalized, UTF-8 safe, consistent JSONL |
| |
|
| | License |
| |
|
| | PLOS articles are released under the PLOS Open Access license, which permits text mining, training, and research reuse. |
| |
|
| | Credits |
| |
|
| | Dataset extraction, cleaning, normalization, restructuring and validation carried out by Zeronex. |
