license: mit
⭐ README — ZERONEX SCIENTIFIC CORPUS (1M CLEAN JSON)
(Made by Zeronex — 2025 Edition)
🚀 Overview
This release contains one of the cleanest scientific corpora ever published. No noise. No XML leftovers. No broken paragraphs. Every file is fully normalized, token-ready, embedding-ready, and AI-training-ready.
All files are professionally structured JSON, signature-stamped, and extracted from scientific metadata with gold-level cleaning rules.
This drop includes:
1️⃣ The MASSIVE 1,000,000 Sample Corpus
A fully cleaned scientific dataset containing:
1,000,000 normalized JSON files
Full abstracts
Metadata
Authors
Categories
Clean paragraphs
Token estimates
Structured fields
Consistent formatting
Unified naming
Signature: "Made by Zeronex"
Every file follows the same perfect schema, with no exceptions.
2️⃣ Sorted Category Files
A second dataset containing category-specific filtered corpora, automatically split by top-level scientific fields:
Examples:
physics.json
quant-ph.json
hep-th.json
astro-ph.json
q-fin.json
etc.
Each file contains only clean, validated entries matching the scientific domain.
This makes it extremely easy to build:
specialized LLMs
domain-specific embeddings
fine-tuned models
retrieval systems
scientific RAG pipelines
3️⃣ Train-Ready Splits
Included:
train.json
valid.json
test.json
Perfectly balanced. No duplicates. No contamination. Totally cleaned. Ready for:
supervised fine-tuning
continued pretraining
embedding model training
scientific QA systems
autoregressive language modeling
These splits are engineered to be plug-and-play for any ML framework (PyTorch, HF Transformers, JAX, etc.).
🔥 Data Quality
This corpus is designed with an industrial-grade cleaning standard, featuring:
✔ No broken text ✔ No incomplete sentences ✔ No XML/HTML noise ✔ No parsing artifacts ✔ Consistent metadata ✔ Robust JSON schema ✔ Perfect normalization ✔ Clean paragraphs ✔ Math extraction support ✔ Subject classification ✔ Zero mixing across domains ✔ Seamless loading at scale ✔ Embedded signature: "Made by Zeronex"
This is not a raw dump. This is not a scraped mess. This is a professional-grade scientific dataset, built to train real models.
⚡ Why This Drop Matters
This dataset can be used immediately for:
LLM pretraining
Domain fine-tuning
RAG systems
Scientific summarization
Research chatbots
Knowledge extraction
Embedding model training
Topic modeling
Graph building
Multi-domain AI assistants
This is the type of corpus used in high-level research labs.
🧬 Signature
Every file includes:
"signature": "Made by Zeronex"
This certifies provenance and protects against dataset plagiarism.
📦 Contents 📁 SampleArXiv_1M/ │── 1M_json_papers/ # main dataset (1,000,000 files) │── categories/ # domain-filtered datasets │── train.json # train split (ready to use) │── valid.json # validation split │── test.json # test split │── README.md # this file
🏁 Final Notes
This is a teaser drop. The full, extended corpus (many millions more) will be released only if this work is shared, supported, and credited properly.
If you use this dataset, please cite the creator:
“Dataset cleaned and structured by Zeronex (2025).”