spencerlima/osaa-metrics

OSAA Metrics

A tidy "One Big Table" of country-year development indicators from the World Bank and UNESCO, paired with a per-indicator metadata table that includes BGE-M3 semantic embeddings for natural-language indicator search.

Built and maintained by Mirian Lima (github.com/mirianlima/osaa-metrics) for the UN Office of the Special Adviser on Africa (OSAA). The companion osaa-metrics Python package consumes this dataset directly via DuckDB's httpfs extension and exposes it through an MCP server, a CLI, and a marimo notebook.

master.parquet — long-format indicator values

One row per (indicator × country × year). ~9.6 M rows, 6 055 indicators, 215 countries, years 1970–2025.

Column	Type	Description
source	string	Data provider (wb, uis)
database	string	Source database (wdi, sdg, opri)
indicator_code	string	Provider-native code (e.g. DT.DOD.DPPG.CD) — unique across the whole table
indicator_name	string	Human-readable indicator name
indicator_description	string	Long-form description
year	int64	Calendar year
value	float64	Indicator value (units vary per indicator — see description)
iso3	string	ISO 3166-1 alpha-3 country code
iso2	string	ISO 3166-1 alpha-2 country code
m49	int64	UN M49 numeric area code
country	string	Country / area name
region	string	UN M49 region
subregion	string	UN M49 sub-region
intermediate_region	string	UN M49 intermediate region (where applicable)
is_ldc	bool	Least Developed Country flag
is_lldc	bool	Landlocked Developing Country flag
is_sids	bool	Small Island Developing State flag
income_group	string	World Bank income classification

meta.parquet — per-indicator metadata + embeddings

One row per indicator (6 055 rows). Used for indicator discovery and semantic search.

Column	Type	Description
source, database, indicator_code	string	Joins to master.parquet
meta_indicator_name	string	Indicator name
meta_indicator_description	string	Indicator description
meta_indicator_type	string	Provider-assigned indicator type
cover_world_pct	float64	% of world countries with at least one observation
cover_region_africa / _americas / _asia / _europe / _oceania	float64	Per-region coverage %
trend_over_years	float64[]	Yearly observation counts as an array (sparkline-ready)
year_start / year_end	int64	First / last year with data
stat_cagr_pct	float64	Compound annual growth rate of the global mean
indicator_text	string	Concatenated text used to compute the embedding
embedding	float32[1024]	BGE-M3 dense embedding of indicator_text

Note: the embedding column is a 1024-dim float array, so rows look very wide in Data Studio. That's expected.

Coverage by source

Source	Database	Indicators
UNESCO Institute for Statistics	SDG	2 506
UNESCO Institute for Statistics	OPRI	2 042
World Bank	WDI	1 507

Quick start

From DuckDB (no auth needed)

INSTALL httpfs; LOAD httpfs;
SELECT * FROM read_parquet('hf://datasets/spencerlima/osaa-metrics/master.parquet') LIMIT 10;
SELECT * FROM read_parquet('hf://datasets/spencerlima/osaa-metrics/meta.parquet')   LIMIT 10;

From Python via the osaa-metrics package

git clone https://github.com/mirianlima/osaa-metrics
cd osaa-metrics
uv sync --extra mcp --extra viz
uv run osaa-mcp   # MCP server for Claude Desktop
# or
uv run marimo edit notebook/main.py   # interactive notebook

Defaults to this dataset (hf://spencerlima/osaa-metrics) — no HF_TOKEN required for read access.

From the datasets library

from datasets import load_dataset
master = load_dataset("spencerlima/osaa-metrics", "master", split="train")
meta   = load_dataset("spencerlima/osaa-metrics", "meta",   split="train")

Semantic indicator search

The embedding column in meta.parquet is a normalized BGE-M3 dense vector of indicator_text. To find indicators by natural-language description:

import polars as pl, numpy as np
from sentence_transformers import SentenceTransformer

meta = pl.read_parquet("hf://datasets/spencerlima/osaa-metrics/meta.parquet")
embs = np.stack(meta["embedding"].to_list())

model = SentenceTransformer("BAAI/bge-m3")
q = model.encode(["youth unemployment in sub-Saharan Africa"], normalize_embeddings=True)
scores = embs @ q.T
top = meta.with_columns(pl.Series("score", scores.flatten())).sort("score", descending=True).head(10)

Licence

Released under CC-BY-4.0. You may share and adapt the data for any purpose, including commercial use, provided you give appropriate credit.

The underlying source data is published by the World Bank and UNESCO Institute for Statistics under their respective open-data terms — please consult the original providers for indicator-level licensing.

Citation

@misc{lima_osaa_metrics_2026,
  author       = {Mirian Lima},
  title        = {OSAA Metrics — Economic \& Trade Indicators},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/spencerlima/osaa-metrics},
  note         = {Companion code: \url{https://github.com/mirianlima/osaa-metrics}}
}

Links

• Source code: github.com/mirianlima/osaa-metrics
• Upstream data: World Bank Open Data, UNESCO UIS
• Embedding model: BAAI/bge-m3