Datasets:
metadata
license: mit
task_categories:
- text-generation
- fill-mask
language:
- en
tags:
- stem
- science
- math
- physics
- biology
- education
pretty_name: Palladium STEM Preview
size_categories:
- 10K<n<100K
⚛️ Palladium-STEM (Preview): High-Density Scientific Corpus
"The Top 0.17% of the Open Web."
Overview
This dataset is a 25,000-document preview of the upcoming Palladium-V2 STEM Corpus. It represents the "Platinum Tier" survivors from a pool of 14.8 million scanned documents, selected for high information density, academic rigor, and reasoning capability.
The "Goldilocks" Methodology
Unlike standard web scrapes, this data was processed using a custom GPU-accelerated refinery:
- CPU Gatekeeper: Filters for academic structure, citations, and mathematical notation (LaTeX, proofs, algorithms).
- GPU Classifier: A BERT-based educational model scores every document on pedagogical quality.
- The Result: A 0.17% yield rate. We rejected 99.83% of the content to leave only high-grade signal.
Data Provenance
- Sources: High-authority domains including
phys.org,wikipedia.org,newscientist.com,betterlesson.com, and university domains (.edu). - Content: Hard sciences (Physics, Chemistry), Mathematics (Proofs, Logic), Computer Science (Algorithms), and Academic Curriculum.
- Quality: Average educational score of 3.78 / 5.0 (vs. standard web average of <1.5).
Usage
This preview is intended for:
- Pre-training LLMs on high-quality reasoning data.
- Fine-tuning models for STEM tasks.
- Evaluating data quality filters.
Full Dataset (Coming Soon)
The full Palladium-V2 corpus (targeting 5M+ documents / 5B+ tokens) is currently mining. [Contact us for early access]