rausch/scientific_corpus_cleaned

Dataset Card for Scientific Corpus (Cleaned)

This corpus contains ≈11 M English scientific documents cleaned via the DataTrove pipeline. It was used to continue pretraining T5-base (EN‑T5-Sci) before sliding-window materialization. Each document is provided as a row in one of 75 Parquet shards together with extensive per-document QA metadata.

Dataset Details

Uses

Direct Use

• Continued pretraining / domain adaptation of encoder-decoder LMs on scientific text.
• Building scientific QA, summarization, or retrieval benchmarks for English.

Dataset Structure

• Split: single train split (≈11 M docs).
• Fields: text (string), id (string), metadata (struct with QA metrics such as length, fasttext score, citation counts, publisher/year).
• Files: 75 Parquet shards + stats/summary/* JSONs with descriptive statistics.

Dataset Creation

Curation Rationale

Provide a reproducible, high-quality English scientific corpus for EN‑T5-Sci pretraining and subsequent cross-lingual transfer.

Source Data

• Data Collection: Unpaywall snapshot curated by the DFKI Scilons team (PDF → text via GROBID).
• Processing: DataTrove + custom scripts (citation removal, structural filtering, FastText EN filter ≥0.75, conservative normalization). Outputs include cleaned text and per-document QA metadata.
• Producers: Scientific publishers indexed by Unpaywall; metadata retains publisher/journal/year when available.

Bias, Risks, and Limitations

-domain bias toward STEM fields; humanities/social sciences underrepresented.

• Potential PII leakage; language filter may drop multilingual documents.
• Residual OCR artifacts may remain.