odoma/gotriple-pretraining-dataset

GoTriple Pretraining dataset

Summary

The GoTriple Pre-training Dataset is a multilingual corpus built from open-access research artefacts harvested via the GoTriple platform. It focuses on Social Sciences and Humanities (SSH) content, addressing their limited presence in standard LLM pre-training corpora.

Current release includes History, Sociology, Environmental Sciences, Psychology and Geography texts (~23.14B tokens).

Intended Use

• Continuous pre-training of LLMs for SSH knowledge.
• Domain adaptation for reasoning, summarisation, QA, and text understanding tasks.
• Multilingual and long-form academic text experimentation.

Not suited for:

• High-stakes factual inference without validation.
• Model training requiring strict licensing confirmation.

Limitations

• PDF extraction inconsistencies.
• Multilingual documents (avg. 1.3--1.5 languages per text).
• Possible near-duplicates.
• Noise from URLs, formulas, encoding issues.

Ethical Considerations

• Dataset filtered for open-access, but users must verify licenses.
• SSH texts may encode cultural or disciplinary biases.
• Language imbalance (European languages dominate).
• Care advised for historically/socially sensitive applications.

Data Composition

• Source: GoTriple API (19.9M artefacts across 27 disciplines).
• Current subset: History, Sociology, Environmental Sciences, Psychology and Geography.
• Languages: 120+ detected; top: EN, DE, PL, FR, ES.
• Token counts: History 4.47B, Sociology 5.33B, Environmental Sciences 5.21B, Psychology 3.65B, Geography 4.46B (total ~23.14B).

Features: raw text, language IDs, spell-check stats, perplexity, PDF metadata, dedupe cluster IDs.

Collection & Processing

1. Metadata filtering (OA + PDF availability)
2. Parallel PDF retrieval (retry + backoff; skip >500MB)
3. PyMuPDF text extraction + heuristics
4. Semantic deduplication (Model2Vec + FAISS, 0.99 similarity)
5. Quality scoring (GlotLID, spell-check, perplexity)

Version History

• v1.0 --- 5 disciplines: History, Sociology, Environmental Sciences, Psychology and Geography (23.14B tokens)
• v0.1 --- History + Sociology (~9.8B tokens)

Credits

This dataset was created by Harshdeep Singh (Odoma) and Matteo Romanello in the context of the EU-funded GRAPHIA project (grant ID: 101188018).