Datasets:

AdhyanshVerma
/

osti

license: gfdl
task_categories:
  - text-generation
language:
  - en
tags:
  - OSTI
  - Scientific
  - Documents
  - Tech
  - Reposrts
  - Conference
  - GNU
pretty_name: osti

DOE OSTI Technical Reports & Conference Papers

558,050 scientific documents from the U.S. Department of Energy's Office of Scientific and Technical Information (OSTI), with lazy full-text pointers to DOE-hosted PDFs.

📊 Dataset Overview

Statistic	Value
Total Records	558,050
Technical Reports	311,823
Conference Papers	246,227
With Full-Text Available	558,050 (100%)
Date Range	1943 – Present
Metadata Size	~348 MB (Parquet)
Estimated PDF Corpus	~80–150 GB
Source	OSTI.GOV API v1

This dataset was harvested using the Lazy Pointer architecture: metadata and full-text URLs are stored locally, while actual PDFs can be fetched on-demand or batched later.

🏗️ Architecture

┌─────────────────────────────────────────────────────────────┐
│  YOUR MACHINE (30GB RAM, 20GB Disk)                         │
│  ┌──────────────────┐    ┌──────────────────┐              │
│  │  Metadata Harvest │───▶│  SQLite / Parquet │              │
│  │  (REST API)       │    │  (348 MB)         │              │
│  └──────────────────┘    └──────────────────┘              │
│           │                                                │
│           ▼                                                │
│  ┌──────────────────┐                                     │
│  │  Lazy Pointers   │  ──▶  Full-Text URLs only          │
│  │  (osti_id, url)  │      No PDFs stored locally          │
│  └──────────────────┘                                     │
└─────────────────────────────────────────────────────────────┘
                           │
                           ▼
┌─────────────────────────────────────────────────────────────┐
│  HUGGING FACE (8TB Storage)                                 │
│  ┌──────────────────┐    ┌──────────────────┐              │
│  │  Metadata        │    │  PDFs (optional) │              │
│  │  osti_metadata   │    │  pdfs/{id}.pdf   │              │
│  │  .parquet        │    │  (streamed)      │              │
│  └──────────────────┘    └──────────────────┘              │
└─────────────────────────────────────────────────────────────┘

Why Lazy Pointers?

Your local machine has 20GB disk — can't hold 100GB+ of PDFs
Your 30GB RAM is used for streaming API responses, not buffering entire datasets
8TB HF storage becomes the actual corpus repository
PDFs are fetched later on a bigger machine, or on-demand for specific records

📁 Files

File	Description	Size
`osti_metadata.parquet`	Core dataset with all metadata + lazy pointers	~348 MB
`osti_metadata.sqlite`	SQLite source (optional, for local querying)	~1 GB

📋 Schema

{
    "osti_id": 944980,                          # Unique OSTI identifier
    "title": "EnergyPlus Analysis Capabilities...",  # Document title
    "authors": '["Author 1", "Author 2"]',       # JSON list of authors
    "abstract": "Worldwide interest in...",      # Abstract / description
    "doi": "10.2172/944980",                   # DOI (if journal article)
    "product_type": "Technical Report",         # Document category
    "publication_date": "2002-09-01",           # ISO-8601 date
    "subjects": '["energy", "building"]',       # JSON list of keywords
    "fulltext_url": "https://www.osti.gov/servlets/purl/944980",  # PDF URL
    "pdf_downloaded": 0,                         # 0 = not yet fetched, 1 = fetched
    "harvested_at": "20260719_163215"           # Harvest batch timestamp
}

Product Types

Type	Count	Full-Text Source
Technical Report	311,823	DOE-hosted (direct PDF) ✅
Conference	246,227	DOE-hosted (direct PDF) ✅

Note: Journal Articles were excluded from this harvest because their full text lives on publisher sites (paywalled), not OSTI servers.

🚀 Quick Start

Load Metadata

import pandas as pd

df = pd.read_parquet("osti/osti_metadata.parquet")
print(f"Total records: {len(df)}")

# Filter by subject
solar = df[df['subjects'].str.contains('solar', case=False, na=False)]
print(f"Solar energy docs: {len(solar)}")

# Get a specific PDF URL
url = df[df['osti_id'] == 944980]['fulltext_url'].values[0]
print(url)  # https://www.osti.gov/servlets/purl/944980

Download a Single PDF

import requests

osti_id = 944980
url = f"https://www.osti.gov/servlets/purl/{osti_id}"

resp = requests.get(url, stream=True)
with open(f"{osti_id}.pdf", "wb") as f:
    for chunk in resp.iter_content(chunk_size=8192):
        f.write(chunk)

Batch Download (for big machines)

# Stream PDFs directly to Hugging Face without touching local disk
from huggingface_hub import HfApi
import requests

api = HfApi()
repo_id = "your-username/osti-pdfs"

for _, row in df.iterrows():
    resp = requests.get(row['fulltext_url'], stream=True)
    if resp.status_code == 200:
        api.upload_file(
            path_or_fileobj=resp.content,
            path_in_repo=f"pdfs/{row['osti_id']}.pdf",
            repo_id=repo_id,
            repo_type="dataset"
        )

🔍 Search & Filter Examples

# By product type
trs = df[df['product_type'] == 'Technical Report']

# By date range
recent = df[df['publication_date'] >= '2020-01-01']

# By keyword in abstract
fusion = df[df['abstract'].str.contains('fusion', case=False, na=False)]

# By subject tag
nuclear = df[df['subjects'].str.contains('nuclear', case=False, na=False)]

# By DOE contract number (in abstract or title)
contract = df[df['abstract'].str.contains('DE-AC02', case=False, na=False)]

🛠️ How This Dataset Was Built

Hardware Constraints

RAM: 30 GB
Local Disk: 20 GB
CPU: 4 v-cores
Remote Storage: 8 TB (Hugging Face)

Harvest Process

Reconnaissance — Queried OSTI API to count records per product_type with has_fulltext=true
Metadata Harvest — Streamed 558K records via REST API (500 records/page), parsed with ijson, stored in SQLite
Checkpointing — Saved last_page after every API request to survive crashes
Compression — Converted SQLite to Parquet with PyArrow row-group chunking to optimize memory
Upload — Pushed Parquet to Hugging Face, freed local disk

Rate Limiting

Conservative: 1 request per second to OSTI servers
Retries with exponential backoff on 429/500 errors
Total harvest time: ~40 minutes for 558K metadata records

📚 OSTI Subject Areas Covered

This dataset spans all 14 DOE subject areas:

Biology and Medicine
Chemistry
Energy Storage, Conversion, and Utilization
Engineering
Environmental Sciences
Fission and Nuclear Technologies
Fossil Fuels
Geosciences
Materials
Mathematics and Computing
National Defense
Physics
Power Generation and Distribution
Renewable Energy

⚠️ Important Notes

Full-text PDFs are NOT included in this repo. Only URLs (lazy pointers) are stored.
PDFs are hosted by the U.S. Department of Energy and are generally public domain or government work.
Some very old records may have scanned PDFs (image-only, no text layer).
Journal Articles with DOIs are excluded — their full text lives on publisher sites.

📜 License & Attribution

Metadata: Harvested from the OSTI.GOV API, a U.S. government service. Metadata is in the public domain.

Full-Text PDFs: Hosted by OSTI. Most DOE-funded research is public domain or available under open-access terms. Verify individual documents for specific licensing.

Dataset Citation:

@dataset{osti_technical_reports_2026,
  title        = {DOE OSTI Technical Reports and Conference Papers (558K Records)},
  author       = {OSTI.GOV},
  year         = 2026,
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/your-username/osti-technical-reports}}
}

🔗 Related Links

OSTI.GOV — Official DOE science search portal
OSTI API Docs — REST API documentation
SciTech Connect (Retired) — Legacy interface
DOE Data Explorer — Related data repository

🤝 Contributing

This is a living dataset. To contribute:

Expand coverage — Add more product types (Patents, Theses, Books)
Enrich metadata — Extract text from PDFs, add embeddings, tag entities
Build downstream datasets — Create domain-specific subsets (e.g., only renewable energy)
Report issues — Open an issue if you find broken full-text URLs

Built with 16GB RAM, 1TB disk, 12 v-cores, and 8TB of Hugging Face storage. No scraping — just polite API harvesting. 🚀