Update README

README.md

@@ -0,0 +1,1051 @@
+---
+language:
+- en
+license: cc0-1.0
+task_categories:
+- feature-extraction
+- text-classification
+- question-answering
+pretty_name: OpenAlex - Complete Academic Research Database
+size_categories:
+- 100M<n<1B
+source_datasets:
+- openalex
+tags:
+- academic
+- research
+- scholarly
+- citations
+- science
+- open-access
+- parquet
+- bibliometrics
+- scientometrics
+dataset_info:
+- config_name: topics
+  features:
+  - name: id
+    dtype: string
+  - name: display_name
+    dtype: string
+  - name: description
+    dtype: string
+  - name: keywords
+    dtype: string
+  - name: subfield_id
+    dtype: string
+  - name: subfield_name
+    dtype: string
+  - name: field_id
+    dtype: string
+  - name: field_name
+    dtype: string
+  - name: domain_id
+    dtype: string
+  - name: domain_name
+    dtype: string
+  - name: siblings
+    dtype: string
+  - name: works_count
+    dtype: int32
+  - name: cited_by_count
+    dtype: int32
+  - name: ids
+    dtype: string
+  - name: created_date
+    dtype: string
+  - name: updated_date
+    dtype: string
+- config_name: publishers
+  features:
+  - name: id
+    dtype: string
+  - name: display_name
+    dtype: string
+  - name: alternate_titles
+    dtype: string
+  - name: hierarchy_level
+    dtype: int32
+  - name: parent_publisher
+    dtype: string
+  - name: country_codes
+    dtype: string
+  - name: homepage_url
+    dtype: string
+  - name: works_count
+    dtype: int32
+  - name: cited_by_count
+    dtype: int32
+  - name: h_index
+    dtype: int32
+  - name: i10_index
+    dtype: int32
+  - name: lineage
+    dtype: string
+  - name: roles
+    dtype: string
+  - name: counts_by_year
+    dtype: string
+  - name: ids
+    dtype: string
+  - name: created_date
+    dtype: string
+  - name: updated_date
+    dtype: string
+- config_name: funders
+  features:
+  - name: id
+    dtype: string
+  - name: display_name
+    dtype: string
+  - name: alternate_titles
+    dtype: string
+  - name: country_code
+    dtype: string
+  - name: description
+    dtype: string
+  - name: homepage_url
+    dtype: string
+  - name: works_count
+    dtype: int32
+  - name: cited_by_count
+    dtype: int32
+  - name: awards_count
+    dtype: int32
+  - name: h_index
+    dtype: int32
+  - name: i10_index
+    dtype: int32
+  - name: roles
+    dtype: string
+  - name: counts_by_year
+    dtype: string
+  - name: ids
+    dtype: string
+  - name: created_date
+    dtype: string
+  - name: updated_date
+    dtype: string
+- config_name: sources
+  features:
+  - name: id
+    dtype: string
+  - name: issn_l
+    dtype: string
+  - name: issn
+    dtype: string
+  - name: display_name
+    dtype: string
+  - name: type
+    dtype: string
+  - name: host_organization
+    dtype: string
+  - name: host_organization_name
+    dtype: string
+  - name: works_count
+    dtype: int32
+  - name: cited_by_count
+    dtype: int32
+  - name: is_oa
+    dtype: bool
+  - name: is_in_doaj
+    dtype: bool
+  - name: is_core
+    dtype: bool
+  - name: homepage_url
+    dtype: string
+  - name: country_code
+    dtype: string
+  - name: h_index
+    dtype: int32
+  - name: i10_index
+    dtype: int32
+  - name: apc_usd
+    dtype: int32
+  - name: alternate_titles
+    dtype: string
+  - name: topics
+    dtype: string
+  - name: counts_by_year
+    dtype: string
+  - name: ids
+    dtype: string
+  - name: created_date
+    dtype: string
+  - name: updated_date
+    dtype: string
+- config_name: institutions
+  features:
+  - name: id
+    dtype: string
+  - name: ror
+    dtype: string
+  - name: display_name
+    dtype: string
+  - name: type
+    dtype: string
+  - name: country_code
+    dtype: string
+  - name: homepage_url
+    dtype: string
+  - name: image_url
+    dtype: string
+  - name: works_count
+    dtype: int32
+  - name: cited_by_count
+    dtype: int32
+  - name: h_index
+    dtype: int32
+  - name: i10_index
+    dtype: int32
+  - name: geo_city
+    dtype: string
+  - name: geo_region
+    dtype: string
+  - name: geo_country
+    dtype: string
+  - name: geo_latitude
+    dtype: float64
+  - name: geo_longitude
+    dtype: float64
+  - name: associated_institutions
+    dtype: string
+  - name: lineage
+    dtype: string
+  - name: topics
+    dtype: string
+  - name: counts_by_year
+    dtype: string
+  - name: roles
+    dtype: string
+  - name: ids
+    dtype: string
+  - name: created_date
+    dtype: string
+  - name: updated_date
+    dtype: string
+- config_name: authors
+  features:
+  - name: id
+    dtype: string
+  - name: orcid
+    dtype: string
+  - name: display_name
+    dtype: string
+  - name: display_name_alternatives
+    dtype: string
+  - name: works_count
+    dtype: int32
+  - name: cited_by_count
+    dtype: int32
+  - name: h_index
+    dtype: int32
+  - name: i10_index
+    dtype: int32
+  - name: two_yr_mean_citedness
+    dtype: float64
+  - name: affiliations
+    dtype: string
+  - name: last_known_institutions
+    dtype: string
+  - name: topics
+    dtype: string
+  - name: topic_share
+    dtype: string
+  - name: counts_by_year
+    dtype: string
+  - name: ids
+    dtype: string
+  - name: created_date
+    dtype: string
+  - name: updated_date
+    dtype: string
+- config_name: works
+  features:
+  - name: id
+    dtype: string
+  - name: doi
+    dtype: string
+  - name: title
+    dtype: string
+  - name: publication_year
+    dtype: int32
+  - name: publication_date
+    dtype: string
+  - name: type
+    dtype: string
+  - name: language
+    dtype: string
+  - name: is_retracted
+    dtype: bool
+  - name: is_paratext
+    dtype: bool
+  - name: cited_by_count
+    dtype: int32
+  - name: fwci
+    dtype: float64
+  - name: referenced_works_count
+    dtype: int32
+  - name: authors_count
+    dtype: int32
+  - name: locations_count
+    dtype: int32
+  - name: is_oa
+    dtype: bool
+  - name: oa_status
+    dtype: string
+  - name: oa_url
+    dtype: string
+  - name: primary_location
+    dtype: string
+  - name: best_oa_location
+    dtype: string
+  - name: locations
+    dtype: string
+  - name: authorships
+    dtype: string
+  - name: biblio_volume
+    dtype: string
+  - name: biblio_issue
+    dtype: string
+  - name: biblio_first_page
+    dtype: string
+  - name: biblio_last_page
+    dtype: string
+  - name: primary_topic
+    dtype: string
+  - name: topics
+    dtype: string
+  - name: keywords
+    dtype: string
+  - name: referenced_works
+    dtype: string
+  - name: related_works
+    dtype: string
+  - name: abstract_inverted_index
+    dtype: string
+  - name: ids
+    dtype: string
+  - name: counts_by_year
+    dtype: string
+  - name: sustainable_development_goals
+    dtype: string
+  - name: indexed_in
+    dtype: string
+  - name: created_date
+    dtype: string
+  - name: updated_date
+    dtype: string
+configs:
+- config_name: topics
+  data_files: "data/topics/*.parquet"
+- config_name: publishers
+  data_files: "data/publishers/*.parquet"
+- config_name: funders
+  data_files: "data/funders/*.parquet"
+- config_name: sources
+  data_files: "data/sources/*.parquet"
+- config_name: institutions
+  data_files: "data/institutions/*.parquet"
+- config_name: authors
+  data_files: "data/authors/*.parquet"
+- config_name: works
+  data_files: "data/works/*.parquet"
+---
+
+# OpenAlex - Complete Academic Research Database
+
+> The world's scholarly research catalog — 449.9K records across 7 entity types, converted to analysis-ready Parquet
+
+## Table of Contents
+
+- [What is it?](#what-is-it)
+- [What is being released?](#what-is-being-released)
+- [How to download and use this dataset](#how-to-download-and-use-this-dataset)
+- [Entity overview](#entity-overview)
+- [Entity relationships](#entity-relationships)
+- [Schema details](#schema-details)
+- [Abstract reconstruction](#abstract-reconstruction)
+- [How it works](#how-it-works)
+- [Dataset card](#dataset-card-for-openalex)
+  - [Dataset summary](#dataset-summary)
+  - [Dataset structure](#dataset-structure)
+  - [Dataset creation](#dataset-creation)
+  - [Considerations for using the data](#considerations-for-using-the-data)
+- [Attribution](#attribution)
+- [Additional information](#additional-information)
+
+## What is it?
+
+[OpenAlex](https://openalex.org) is a free and open catalog of the world's scholarly research system — papers, authors, institutions, journals, topics, publishers, and funders — maintained by [OurResearch](https://ourresearch.org/). It is the open replacement for the discontinued Microsoft Academic Graph (MAG) and currently indexes over 250 million scholarly works with their full citation networks, authorship chains, institutional affiliations, and topic classifications.
+
+This dataset is a complete conversion of the [OpenAlex snapshot](https://docs.openalex.org/download-all-data/openalex-snapshot) from its native gzipped JSON Lines format into sharded, ZSTD-compressed Parquet files. The snapshot date is **2026-04** and contains **449.9K total records** across 7 entity types. The data is stored as one Parquet file per million rows, making it straightforward to query with DuckDB, load with the `datasets` library, or process with any tool that reads Parquet.
+
+We believe this is one of the most complete and accessible mirrors of OpenAlex data available on Hugging Face. The Parquet format enables direct SQL queries via DuckDB's `hf://` protocol without downloading anything first.
+
+## What is being released?
+
+The dataset is organized as sharded Parquet files per entity type. Each entity type is a separate HuggingFace dataset configuration, so you can load just the entities you need.
+
+```
+data/
+  works/
+    works-00000.parquet      scholarly works (~1M rows each)
+    works-00001.parquet
+    ...
+  authors/
+    authors-00000.parquet    researchers and their metrics
+    ...
+  sources/
+    sources-00000.parquet    journals, repositories, conferences
+    ...
+  institutions/
+    institutions-00000.parquet  universities, labs, companies
+    ...
+  topics/
+    topics-00000.parquet     research topic taxonomy
+  publishers/
+    publishers-00000.parquet academic publishers
+    ...
+  funders/
+    funders-00000.parquet    funding organizations
+    ...
+```
+
+Each shard contains up to 1 million rows, compressed with Zstandard for a good balance of size and query speed. Nested and complex fields (authorships, locations, topics, etc.) are stored as JSON strings, queryable with DuckDB's `json_extract()` or Python's `json.loads()`.
+
+## How to download and use this dataset
+
+You can load the full dataset, a single entity type, or even query across entities with joins. The dataset uses the standard Hugging Face Parquet layout, so it works out of the box with DuckDB, the `datasets` library, `pandas`, and `huggingface_hub`.
+
+### Using DuckDB
+
+DuckDB can read Parquet files directly from Hugging Face without downloading anything first. This is the fastest way to explore the data:
+
+```sql
+-- Most-cited works of all time
+SELECT id, title, publication_year, cited_by_count, doi, oa_status
+FROM 'hf://datasets/open-index/open-alex/data/works/*.parquet'
+WHERE cited_by_count > 1000
+ORDER BY cited_by_count DESC
+LIMIT 20;
+```
+
+```sql
+-- Top authors by h-index
+SELECT id, display_name, h_index, i10_index, works_count, cited_by_count
+FROM 'hf://datasets/open-index/open-alex/data/authors/*.parquet'
+ORDER BY h_index DESC
+LIMIT 20;
+```
+
+```sql
+-- Open access rates by year
+SELECT publication_year,
+       COUNT(*) as total,
+       SUM(CASE WHEN is_oa THEN 1 ELSE 0 END) as oa_count,
+       ROUND(100.0 * SUM(CASE WHEN is_oa THEN 1 ELSE 0 END) / COUNT(*), 1) as oa_pct
+FROM 'hf://datasets/open-index/open-alex/data/works/*.parquet'
+WHERE publication_year BETWEEN 2000 AND 2025
+GROUP BY publication_year
+ORDER BY publication_year;
+```
+
+```sql
+-- Top US institutions by research output
+SELECT display_name, type, geo_city, works_count, cited_by_count, h_index
+FROM 'hf://datasets/open-index/open-alex/data/institutions/*.parquet'
+WHERE country_code = 'US'
+ORDER BY works_count DESC
+LIMIT 20;
+```
+
+```sql
+-- Most common research topics
+SELECT id, display_name, subfield_name, field_name, domain_name, works_count
+FROM 'hf://datasets/open-index/open-alex/data/topics/*.parquet'
+ORDER BY works_count DESC
+LIMIT 20;
+```
+
+```sql
+-- Extract author affiliations from nested JSON
+SELECT id, display_name,
+       json_extract_string(last_known_institutions, '$[0].display_name') as institution,
+       json_extract_string(last_known_institutions, '$[0].country_code') as country
+FROM 'hf://datasets/open-index/open-alex/data/authors/*.parquet'
+WHERE last_known_institutions IS NOT NULL
+ORDER BY h_index DESC
+LIMIT 20;
+```
+
+```sql
+-- Join works to authors via the authorships JSON
+SELECT w.title, w.publication_year, w.cited_by_count, a.display_name, a.h_index
+FROM 'hf://datasets/open-index/open-alex/data/works/*.parquet' w,
+     'hf://datasets/open-index/open-alex/data/authors/*.parquet' a
+WHERE w.cited_by_count > 5000
+  AND a.id = json_extract_string(w.authorships, '$[0].author.id')
+ORDER BY w.cited_by_count DESC
+LIMIT 20;
+```
+
+```sql
+-- Largest publishers by work count
+SELECT id, display_name, works_count, cited_by_count, h_index,
+       json_extract_string(country_codes, '$[0]') as country
+FROM 'hf://datasets/open-index/open-alex/data/publishers/*.parquet'
+ORDER BY works_count DESC
+LIMIT 20;
+```
+
+### Using `datasets`
+
+```python
+from datasets import load_dataset
+
+# Stream works without downloading everything first
+ds = load_dataset("open-index/open-alex", "works", split="train", streaming=True)
+for work in ds:
+    print(work["id"], work["title"], work["cited_by_count"])
+
+# Load a single entity type into memory
+authors = load_dataset("open-index/open-alex", "authors", split="train")
+print(f"{len(authors):,} authors loaded")
+
+# Load smaller entities fully (topics, publishers, funders fit in memory easily)
+topics = load_dataset("open-index/open-alex", "topics", split="train")
+```
+
+### Using `huggingface_hub`
+
+```python
+from huggingface_hub import snapshot_download
+
+# Download only authors (~70 GB compressed, ~114M rows)
+snapshot_download(
+    "open-index/open-alex",
+    repo_type="dataset",
+    local_dir="./openalex/",
+    allow_patterns="data/authors/*",
+)
+
+# Download small entity types only (~500 MB total)
+snapshot_download(
+    "open-index/open-alex",
+    repo_type="dataset",
+    local_dir="./openalex/",
+    allow_patterns=["data/topics/*", "data/publishers/*", "data/funders/*",
+                     "data/sources/*", "data/institutions/*"],
+)
+```
+
+For faster downloads, install `pip install huggingface_hub[hf_transfer]` and set `HF_HUB_ENABLE_HF_TRANSFER=1`.
+
+### Using pandas + DuckDB
+
+```python
+import duckdb
+
+conn = duckdb.connect()
+
+# Citation distribution: what does a "typical" paper look like?
+df = conn.sql("""
+    SELECT
+        percentile_disc(0.50) WITHIN GROUP (ORDER BY cited_by_count) AS p50,
+        percentile_disc(0.90) WITHIN GROUP (ORDER BY cited_by_count) AS p90,
+        percentile_disc(0.99) WITHIN GROUP (ORDER BY cited_by_count) AS p99,
+        percentile_disc(0.999) WITHIN GROUP (ORDER BY cited_by_count) AS p999,
+        AVG(cited_by_count) AS mean
+    FROM read_parquet('hf://datasets/open-index/open-alex/data/works/*.parquet')
+""").df()
+print(df)
+```
+
+## Entity overview
+
+| Entity | Records | Description |
+|---|---|---|
+| **Topics** | 4.5K | Research topics with hierarchical classification (domain → field → subfield → topic) |
+| **Publishers** | 10.7K | Academic publishers with hierarchy levels and country information |
+| **Funders** | 32.4K | Research funding organizations with award counts and cross-references |
+| **Sources** | 280.7K | Journals, repositories, conferences, and ebook platforms with ISSN, DOAJ status, and APC pricing |
+| **Institutions** | 121.5K | Universities, research centers, companies, and government bodies with ROR IDs and geolocation |
+| **Authors** | n/a | Researchers with ORCID IDs, h-index, affiliations, and publication statistics |
+| **Works** | n/a | Scholarly works (articles, books, datasets) with citations, DOIs, topics, authorships, and open access status |
+
+## Entity relationships
+
+OpenAlex models academic research as an interconnected graph. Works are the central entity, linked to authors via authorships, to journals and repositories via locations, and to each other via citation networks. The topic hierarchy (domain > field > subfield > topic) provides a four-level classification for every work and author.
+
+```
+                          ┌──────────────┐
+                          │    Works     │  ← central entity (~492M)
+                          └──────┬───────┘
+                 ┌───────────────┼───────────────────┐
+                 │               │                   │
+          ┌──────▼──────┐ ┌─────▼──────┐  ┌─────────▼────────┐
+          │ Authorships │ │  Locations  │  │ Referenced Works  │
+          │  (nested)   │ │  (nested)   │  │   (citations)    │
+          └──────┬──────┘ └─────┬──────┘  └──────────────────┘
+                 │              │
+          ┌──────▼──────┐ ┌────▼─────┐
+          │   Authors   │ │  Sources │  journals, repos, conferences
+          │   (~114M)   │ │  (~281K) │
+          └──────┬──────┘ └────┬─────┘
+                 │              │
+      ┌──────────▼──────┐  ┌───▼────────┐
+      │  Institutions   │  │ Publishers │
+      │    (~122K)      │  │   (~11K)   │
+      └─────────────────┘  └────────────┘
+
+  Topics (~4.5K) ── Subfields ── Fields ── Domains
+  Funders (~32K) ── awards ── Works
+```
+
+## Schema details
+
+### Topics
+
+Research topics with hierarchical classification (domain → field → subfield → topic).
+
+| Column | Type | Description |
+|---|---|---|
+| `id` | string | |
+| `display_name` | string | |
+| `description` | string | |
+| `keywords` | string | |
+| `subfield_id` | string | |
+| `subfield_name` | string | |
+| `field_id` | string | |
+| `field_name` | string | |
+| `domain_id` | string | |
+| `domain_name` | string | |
+| `siblings` | string | |
+| `works_count` | int32 | |
+| `cited_by_count` | int32 | |
+| `ids` | string | |
+| `created_date` | string | |
+| `updated_date` | string | |
+
+### Publishers
+
+Academic publishers with hierarchy levels and country information.
+
+| Column | Type | Description |
+|---|---|---|
+| `id` | string | |
+| `display_name` | string | |
+| `alternate_titles` | string | |
+| `hierarchy_level` | int32 | |
+| `parent_publisher` | string | |
+| `country_codes` | string | |
+| `homepage_url` | string | |
+| `works_count` | int32 | |
+| `cited_by_count` | int32 | |
+| `h_index` | int32 | |
+| `i10_index` | int32 | |
+| `lineage` | string | |
+| `roles` | string | |
+| `counts_by_year` | string | |
+| `ids` | string | |
+| `created_date` | string | |
+| `updated_date` | string | |
+
+#### Data completeness
+
+Fields below 100% population:
+
+| Field | Population | Est. Count |
+|---|---|---|
+| `alternate_titles` | 10.0% | 1.1K |
+| `parent_publisher` | 0.0% | 0 |
+| `country_codes` | 90.0% | 9.6K |
+| `homepage_url` | 80.0% | 8.6K |
+| `counts_by_year` | 90.0% | 9.6K |
+
+### Funders
+
+Research funding organizations with award counts and cross-references.
+
+| Column | Type | Description |
+|---|---|---|
+| `id` | string | |
+| `display_name` | string | |
+| `alternate_titles` | string | |
+| `country_code` | string | |
+| `description` | string | |
+| `homepage_url` | string | |
+| `works_count` | int32 | |
+| `cited_by_count` | int32 | |
+| `awards_count` | int32 | |
+| `h_index` | int32 | |
+| `i10_index` | int32 | |
+| `roles` | string | |
+| `counts_by_year` | string | |
+| `ids` | string | |
+| `created_date` | string | |
+| `updated_date` | string | |
+
+#### Data completeness
+
+Fields below 100% population:
+
+| Field | Population | Est. Count |
+|---|---|---|
+| `alternate_titles` | 87.5% | 28.4K |
+| `description` | 56.2% | 18.2K |
+| `homepage_url` | 53.1% | 17.2K |
+
+### Sources
+
+Journals, repositories, conferences, and ebook platforms with ISSN, DOAJ status, and APC pricing.
+
+| Column | Type | Description |
+|---|---|---|
+| `id` | string | |
+| `issn_l` | string | |
+| `issn` | string | |
+| `display_name` | string | |
+| `type` | string | |
+| `host_organization` | string | |
+| `host_organization_name` | string | |
+| `works_count` | int32 | |
+| `cited_by_count` | int32 | |
+| `is_oa` | bool | |
+| `is_in_doaj` | bool | |
+| `is_core` | bool | |
+| `homepage_url` | string | |
+| `country_code` | string | |
+| `h_index` | int32 | |
+| `i10_index` | int32 | |
+| `apc_usd` | int32 | |
+| `alternate_titles` | string | |
+| `topics` | string | |
+| `counts_by_year` | string | |
+| `ids` | string | |
+| `created_date` | string | |
+| `updated_date` | string | |
+
+#### Data completeness
+
+Fields below 100% population:
+
+| Field | Population | Est. Count |
+|---|---|---|
+| `issn_l` | 61.4% | 172.4K |
+| `issn` | 61.4% | 172.4K |
+| `host_organization` | 25.7% | 72.2K |
+| `host_organization_name` | 25.4% | 71.2K |
+| `homepage_url` | 26.4% | 74.2K |
+| `country_code` | 42.5% | 119.3K |
+| `apc_usd` | 3.2% | 9.0K |
+| `alternate_titles` | 23.2% | 65.2K |
+| `topics` | 92.9% | 260.6K |
+| `counts_by_year` | 93.2% | 261.6K |
+| `created_date` | 93.2% | 261.6K |
+
+### Institutions
+
+Universities, research centers, companies, and government bodies with ROR IDs and geolocation.
+
+| Column | Type | Description |
+|---|---|---|
+| `id` | string | |
+| `ror` | string | |
+| `display_name` | string | |
+| `type` | string | |
+| `country_code` | string | |
+| `homepage_url` | string | |
+| `image_url` | string | |
+| `works_count` | int32 | |
+| `cited_by_count` | int32 | |
+| `h_index` | int32 | |
+| `i10_index` | int32 | |
+| `geo_city` | string | |
+| `geo_region` | string | |
+| `geo_country` | string | |
+| `geo_latitude` | float64 | |
+| `geo_longitude` | float64 | |
+| `associated_institutions` | string | |
+| `lineage` | string | |
+| `topics` | string | |
+| `counts_by_year` | string | |
+| `roles` | string | |
+| `ids` | string | |
+| `created_date` | string | |
+| `updated_date` | string | |
+
+#### Data completeness
+
+Fields below 100% population:
+
+| Field | Population | Est. Count |
+|---|---|---|
+| `country_code` | 94.2% | 114.5K |
+| `homepage_url` | 99.2% | 120.5K |
+| `image_url` | 10.7% | 13.1K |
+| `geo_region` | 38.0% | 46.2K |
+| `associated_institutions` | 32.2% | 39.2K |
+| `topics` | 88.4% | 107.5K |
+| `counts_by_year` | 86.8% | 105.4K |
+
+### Authors
+
+Researchers with ORCID IDs, h-index, affiliations, and publication statistics.
+
+| Column | Type | Description |
+|---|---|---|
+| `id` | string | |
+| `orcid` | string | |
+| `display_name` | string | |
+| `display_name_alternatives` | string | |
+| `works_count` | int32 | |
+| `cited_by_count` | int32 | |
+| `h_index` | int32 | |
+| `i10_index` | int32 | |
+| `two_yr_mean_citedness` | float64 | |
+| `affiliations` | string | |
+| `last_known_institutions` | string | |
+| `topics` | string | |
+| `topic_share` | string | |
+| `counts_by_year` | string | |
+| `ids` | string | |
+| `created_date` | string | |
+| `updated_date` | string | |
+
+### Works
+
+Scholarly works (articles, books, datasets) with citations, DOIs, topics, authorships, and open access status.
+
+| Column | Type | Description |
+|---|---|---|
+| `id` | string | |
+| `doi` | string | |
+| `title` | string | |
+| `publication_year` | int32 | |
+| `publication_date` | string | |
+| `type` | string | |
+| `language` | string | |
+| `is_retracted` | bool | |
+| `is_paratext` | bool | |
+| `cited_by_count` | int32 | |
+| `fwci` | float64 | |
+| `referenced_works_count` | int32 | |
+| `authors_count` | int32 | |
+| `locations_count` | int32 | |
+| `is_oa` | bool | |
+| `oa_status` | string | |
+| `oa_url` | string | |
+| `primary_location` | string | |
+| `best_oa_location` | string | |
+| `locations` | string | |
+| `authorships` | string | |
+| `biblio_volume` | string | |
+| `biblio_issue` | string | |
+| `biblio_first_page` | string | |
+| `biblio_last_page` | string | |
+| `primary_topic` | string | |
+| `topics` | string | |
+| `keywords` | string | |
+| `referenced_works` | string | |
+| `related_works` | string | |
+| `abstract_inverted_index` | string | |
+| `ids` | string | |
+| `counts_by_year` | string | |
+| `sustainable_development_goals` | string | |
+| `indexed_in` | string | |
+| `created_date` | string | |
+| `updated_date` | string | |
+
+
+## Abstract reconstruction
+
+The `abstract_inverted_index` field on works stores abstracts as an inverted index for space efficiency. This is OpenAlex's native format — each word maps to an array of positions where it appears in the abstract. To reconstruct the full text:
+
+### Python
+
+```python
+import json
+
+def reconstruct_abstract(inverted_index_json):
+    """Reconstruct abstract text from OpenAlex inverted index format."""
+    if not inverted_index_json:
+        return None
+    idx = json.loads(inverted_index_json)
+    words = []
+    for word, positions in idx.items():
+        for pos in positions:
+            words.append((pos, word))
+    words.sort()
+    return " ".join(w for _, w in words)
+
+# Example usage with DuckDB
+import duckdb
+conn = duckdb.connect()
+df = conn.sql("""
+    SELECT id, title, abstract_inverted_index
+    FROM read_parquet('hf://datasets/open-index/open-alex/data/works/*.parquet')
+    WHERE abstract_inverted_index IS NOT NULL
+    LIMIT 5
+""").df()
+
+df["abstract"] = df["abstract_inverted_index"].apply(reconstruct_abstract)
+print(df[["id", "title", "abstract"]].to_string())
+```
+
+### DuckDB
+
+```sql
+-- Count works with abstracts by year
+SELECT publication_year, COUNT(*) as total,
+       SUM(CASE WHEN abstract_inverted_index IS NOT NULL THEN 1 ELSE 0 END) as with_abstract,
+       ROUND(100.0 * SUM(CASE WHEN abstract_inverted_index IS NOT NULL THEN 1 ELSE 0 END) / COUNT(*), 1) as pct
+FROM 'hf://datasets/open-index/open-alex/data/works/*.parquet'
+WHERE publication_year BETWEEN 2000 AND 2025
+GROUP BY publication_year
+ORDER BY publication_year;
+```
+
+## How it works
+
+The pipeline is built in Go and converts the OpenAlex S3 snapshot into sharded Parquet files using the [parquet-go](https://github.com/parquet-go/parquet-go) library.
+
+**Manifest fetch.** For each entity type, the pipeline downloads the Redshift-compatible manifest from `https://openalex.s3.amazonaws.com/data/{entity}/manifest`. The manifest lists every data file with its S3 URL, compressed size, and record count.
+
+**Part download.** Each manifest entry points to a gzip-compressed JSON Lines file on S3 (typically under 2 GB each). Parts are downloaded sequentially per entity with HTTP resume support, preserving the `updated_date=YYYY-MM-DD/` partition structure locally.
+
+**Streaming conversion.** Each `.gz` part is streamed through a gzip decompressor and parsed line-by-line. Fields are extracted with [gjson](https://github.com/tidwall/gjson) (zero-allocation JSON parsing) and written to sharded Parquet files — 1 million rows per shard, 500K rows per row group, ZSTD compression. Complex nested structures (authorships, locations, topics) are preserved as raw JSON strings for maximum flexibility.
+
+**Incremental publishing.** As each entity type finishes conversion, its Parquet shards are committed to Hugging Face immediately via the Hub API. A serial commit queue prevents conflicts. Downloaded parts are deleted as soon as conversion completes, and Parquet shards are deleted after successful upload — this keeps disk usage minimal even for the 600+ GB works entity.
+
+**Quality tracking.** Every 1000th row is sampled during conversion to measure field population rates. These statistics drive the data completeness tables in this README.
+
+# Dataset card for OpenAlex
+
+## Dataset summary
+
+This dataset is a complete mirror of the [OpenAlex](https://openalex.org) academic research database, converted from gzipped JSON Lines to sharded Parquet. OpenAlex is maintained by [OurResearch](https://ourresearch.org/) as the open replacement for the discontinued Microsoft Academic Graph (MAG). It indexes scholarly works, their authors, the journals and repositories where they're published, the institutions where research happens, the topics they cover, and the organizations that fund them.
+
+The dataset is intended for research, analysis, and training. Common use cases include:
+
+- **Bibliometrics and scientometrics** — citation analysis, h-index computation, collaboration networks
+- **Research trend analysis** — tracking topics, open access adoption, and funding patterns over time
+- **Knowledge graph construction** — the entities form a richly interconnected graph via citations, authorships, and affiliations
+- **Language model training** — titles, abstracts (via inverted index), and topic classifications
+- **Institution and funder analysis** — research output by country, university rankings, funding effectiveness
+- **Open access monitoring** — tracking the proportion and type of open access across journals and institutions
+
+## Dataset structure
+
+### Data instances
+
+Here is an example work record showing the flat Parquet schema with nested fields as JSON strings:
+
+```json
+{
+  "id": "https://openalex.org/W2741809807",
+  "doi": "https://doi.org/10.1038/s41586-019-1099-1",
+  "title": "Quantum supremacy using a programmable superconducting processor",
+  "publication_year": 2019,
+  "type": "article",
+  "language": "en",
+  "is_retracted": false,
+  "cited_by_count": 4521,
+  "is_oa": true,
+  "oa_status": "green",
+  "authorships": "[{\"author_position\":\"first\",\"author\":{\"id\":\"https://openalex.org/A5003442464\",\"display_name\":\"Frank Arute\"},\"institutions\":[{\"id\":\"https://openalex.org/I4210164761\",\"display_name\":\"Google\"}]}]",
+  "primary_topic": "{\"id\":\"https://openalex.org/T10001\",\"display_name\":\"Quantum Computing Algorithms and Complexity\",\"subfield\":{\"display_name\":\"Computational Theory and Mathematics\"}}",
+  "abstract_inverted_index": "{\"The\":[0],\"promise\":[1],\"of\":[2,7],\"quantum\":[3],\"computers\":[4],...}",
+  "updated_date": "2026-03-15T00:00:00"
+}
+```
+
+### Data splits
+
+Each entity type is a separate HuggingFace dataset configuration. You can load any combination:
+
+```python
+from datasets import load_dataset
+
+# Load a single entity type
+works = load_dataset("open-index/open-alex", "works", split="train", streaming=True)
+authors = load_dataset("open-index/open-alex", "authors", split="train")
+topics = load_dataset("open-index/open-alex", "topics", split="train")
+```
+
+## Dataset creation
+
+### Curation rationale
+
+OpenAlex provides the most comprehensive open catalog of scholarly research, but the native snapshot format (330 GB of gzipped JSON Lines partitioned by update date) is not directly queryable. Converting to Parquet unlocks SQL-based exploration via DuckDB, streaming with the `datasets` library, and efficient columnar scans — all without requiring users to download and decompress hundreds of gigabytes of raw data.
+
+By publishing on Hugging Face, the data becomes immediately queryable with DuckDB (via `hf://` paths), streamable with the `datasets` library, and downloadable in bulk or by entity type.
+
+### Source data
+
+All data is sourced from the [OpenAlex S3 snapshot](https://docs.openalex.org/download-all-data/openalex-snapshot), a public S3 bucket maintained by OurResearch. The snapshot is updated monthly and contains the complete OpenAlex database as gzip-compressed JSON Lines files partitioned by `updated_date`.
+
+### Data processing steps
+
+1. **Manifest download.** The pipeline fetches the Redshift-compatible manifest for each entity type from the OpenAlex S3 bucket. Each manifest lists all data files with their URLs, compressed sizes, and record counts.
+
+2. **Part download.** Each data file (~2 GB compressed) is downloaded from S3 via HTTPS with resume support. Files preserve the `updated_date` partition structure locally.
+
+3. **Streaming conversion.** Each gzipped part is streamed through a decompressor and parsed line-by-line. Fields are extracted using zero-allocation JSON parsing and written to sharded Parquet files (1M rows per shard, ZSTD compression). Complex nested fields are preserved as raw JSON strings.
+
+4. **Quality tracking.** Every 1000th row is sampled to compute field population statistics. These drive the data completeness tables in this README.
+
+5. **Publishing.** Parquet shards are committed to this Hugging Face repository per entity type. Downloaded parts and converted shards are cleaned up after each successful commit to minimize disk usage.
+
+No filtering, deduplication, or transformation is applied beyond the format conversion. The data reflects exactly what OpenAlex provides in their snapshot.
+
+### Personal and sensitive information
+
+This dataset contains researcher names, institutional affiliations, and ORCID identifiers as they appear in the public OpenAlex database. This information is sourced from published academic papers and is widely available through publisher metadata, Crossref, and other bibliographic databases. No additional PII processing has been applied.
+
+## Considerations for using the data
+
+### Social impact
+
+By providing the complete OpenAlex catalog in an accessible Parquet format, we aim to democratize access to scholarly metadata. Researchers, librarians, policymakers, and developers can analyze global research output, track open access progress, study collaboration patterns, and build tools for scientific discovery — all without needing to set up complex infrastructure.
+
+### Discussion of biases
+
+OpenAlex inherits biases from its data sources (Crossref, PubMed, institutional repositories, etc.). English-language publications and STEM fields are overrepresented. Journals indexed by major databases have better coverage than regional or non-English publications. Citation counts reflect existing power structures in academia and should not be treated as measures of research quality.
+
+The topic classification system is algorithmically assigned and may not perfectly capture interdisciplinary work or emerging fields.
+
+### Known limitations
+
+- **Nested fields are JSON strings.** Fields like `authorships`, `locations`, `topics`, `counts_by_year`, and `ids` are stored as JSON-encoded strings, not native Parquet nested types. Use `json_extract()` in DuckDB or `json.loads()` in Python to parse them.
+- **`abstract_inverted_index` requires reconstruction.** Abstracts are stored as inverted indices (word → position arrays), not plain text. See the [Abstract reconstruction](#abstract-reconstruction) section for code to reconstruct them.
+- **IDs are full URLs.** OpenAlex IDs are stored as full URLs (e.g., `https://openalex.org/W2741809807`), not short identifiers. Extract the suffix if you need compact IDs.
+- **`updated_date` reflects last modification.** A record's `updated_date` is when OpenAlex last modified it, not when the underlying work was published or cited.
+- **Citation counts are point-in-time.** The `cited_by_count` reflects the snapshot date, not real-time counts.
+- **Some API-only fields are missing.** Fields like `content_urls` on works are only available through the OpenAlex API, not in the snapshot.
+- **`concepts` entity is excluded.** OpenAlex has deprecated concepts in favor of topics. This dataset includes topics but not the legacy concepts entity.
+
+## Attribution
+
+This dataset is derived from [OpenAlex](https://openalex.org), a free and open catalog of the world's scholarly research system. OpenAlex is maintained by [OurResearch](https://ourresearch.org/) and indexes over 250 million scholarly works with their authors, institutions, journals, topics, publishers, and funders. All OpenAlex data is released under the [CC0 1.0 Universal (Public Domain Dedication)](https://creativecommons.org/publicdomain/zero/1.0/).
+
+If you use this data in research, please cite OpenAlex:
+
+> Priem, J., Piwowar, H., & Orr, R. (2022). OpenAlex: A fully-open index of scholarly works, authors, venues, institutions, and concepts. *ArXiv*. https://arxiv.org/abs/2205.01833
+
+For more information about citing OpenAlex, see: [https://docs.openalex.org/how-to-use-the-api/get-started](https://docs.openalex.org/how-to-use-the-api/get-started)
+
+## Additional information
+
+### Licensing
+
+The dataset is released under **CC0 1.0 Universal (Public Domain)**, the same license as the underlying OpenAlex data. You are free to use, modify, and redistribute the data for any purpose without attribution, though citation is appreciated.
+
+### Contact
+
+For questions, feedback, or issues with this Parquet conversion, please open a discussion on the [Community tab](https://huggingface.co/datasets/open-index/open-alex/discussions).
+
+For questions about the underlying data, see the [OpenAlex documentation](https://docs.openalex.org/) or contact the OpenAlex team at [support@openalex.org](mailto:support@openalex.org).
+
+*Snapshot: 2026-04*