Themelios-11

Version: 1.0
Size: 11.7~ GB (~3.5 B tokens)
Curated by: CreatorDevX License: Mixed (Non-Commercial, Research-Only)
Primary language: English
Intended use: Foundation-level pretraining for reasoning, dialogue, and multi-agent cognition research.

---

📘 Overview

Themelios-11 (Greek: θεμέλιος, “foundation”) is a high-quality, multi-domain pretraining corpus curated for research. Unlike typical large scrapes, it emphasizes semantic density, domain diversity, and internal coherence over raw token count.

The dataset merges public and open research texts, structured factual knowledge, academic material, and diverse conversational sources to create a compact but expressive training base suitable for mid-scale LLM pretraining or advanced adaptive finetuning.

---

Composition

File	Size (GB)	Domain	Source / License
books.txt	2.61	Long-form narrative and expository text	Public Domain (Project Gutenberg / similar)
wikidata5m_text.txt	2.28	Structured factual and entity descriptions	CC-BY-SA 3.0
dialogs.txt	2.07	Multi-turn conversational data	Mixed open & synthetic, Non-commercial
arxiv.txt	1.07	Full scientific papers	CC-BY 4.0 / CC-BY-NC-SA 4.0
arxiv_abstracts.txt	0.41	Condensed scientific abstracts	CC-BY 4.0
stackoverflow.txt	0.83	Technical Q&A and code reasoning	CC-BY-SA 4.0
code1.txt	0.65	Programming and algorithmic snippets	CC-BY-SA 4.0
webscale.txt	0.56	General web text, blogs, essays	Non-commercial fair-use selection
100MRedditTrainDataset.txt	0.42	Informal social dialogue	Non-commercial fair use
maths.txt	0.39	Mathematical explanations, symbolic reasoning	CC-BY / Public Domain
simplewiki.txt	0.23	Simplified factual exposition	CC-BY-SA 3.0
stanfordphilosophy.txt	0.14	Philosophical and conceptual essays	CC-BY-NC-SA 3.0
TimeMagazine.txt	0.11	Journalistic and historical writing	Copyrighted / Non-commercial use
Currentaffairs.txt	0.014	Contemporary events summaries	Copyrighted / Non-commercial use

Total: ~11.7 GB raw text across 13 domains.
Each sub-corpus retains its original license and copyright.

---

Structure

All files are plain UTF-8 text, newline-delimited.
Each document or paragraph block is separated by a blank line.
No metadata or markup is required for loading — tokenizers can stream from raw text.

---

Statistics

Metric	Value
Total files	13
Approx. tokens	~5–6 B
Average document length	1.8 k tokens
Primary language	English (98 %)
Domain coverage	Narrative · Scientific · Technical · Philosophical · Conversational

---

Suggested Use

• Pretraining mid-scale language models (≤ 3 B params)
• Adaptive continual-pretraining for reasoning or dialogue agents
• Synthetic data generation and self-play bootstrapping
• Evaluation of multi-domain transfer and cross-context generalization

Do not use for commercial applications or redistribution of raw text.

---

Limitations

• Contains texts under different open and restricted licenses; redistribution of the raw corpus may violate original terms.
• English-only; lacks multilingual coverage.
• Some web or Reddit content may include informal or biased language.
• Temporal information uneven (1980-2025).
• Not filtered for personally identifiable data beyond standard public-domain scrubbing.

---

License Summary

Overall License:

Mixed-source corpus intended for non-commercial research only.
Each file retains its upstream license (CC-BY, CC-BY-SA, CC-BY-NC-SA, or equivalent).
Do not redistribute or use commercially.
Model weights trained on Themelios-11 may be released separately under open-model licenses (e.g., OpenRAIL-M) with attribution.

If you reference this dataset in a publication, cite as:

S.R. (2025). Themelios-11: A High-Quality Multi-Domain Corpus for Research-Grade LLM Pretraining.

---

Citation (BibTeX)

@misc{themelios11,
  author       = {S.R.},
  title        = {Themelios-11: A High-Quality Multi-Domain Corpus for Research-Grade LLM Pretraining},
  year         = {2025},
  howpublished = {Hugging Face Datasets},
  note         = {Version 1.0, non-commercial research use only}
}