Refresh dataset card: 2026-05-21 ML stack + voidly-bench-v1

README.md

@@ -12,18 +12,19 @@ tags:
   - human-rights
   - network-measurement
   - geopolitics
-pretty_name: Voidly Global Censorship Index
+pretty_name: Voidly OONI Censorship Historical
 size_categories:
   - 1M<n<10M
 ---
 
-# Voidly Global Censorship Index
+# Voidly OONI Censorship Historical
 
-**The most comprehensive open dataset for internet censorship research and ML.**
+**A 10-year open archive for internet censorship research and ML.**
 
 ## Dataset Description
 
-This dataset contains 10 years of global internet censorship measurements from 120+ countries, including:
+This dataset contains 10 years of global internet censorship measurements
+from 120+ countries:
 
 - **1.6M+ daily measurements** (2017-2026)
 - **37K detected anomaly spikes**
@@ -38,9 +39,8 @@ This dataset contains 10 years of global internet censorship measurements from 1
 
 | File | Description | Rows |
 |------|-------------|------|
-| `ooni-historical.parquet` | Daily measurements by country/test | 1.6M |
-| `censorship-incidents.parquet` | Labeled anomaly spikes | 37K |
-| `known-events.json` | Major censorship events | 25+ |
+| `data/ooni-historical.parquet` | Daily measurements by country/test | 1.6M |
+| `data/censorship-incidents.parquet` | Labeled anomaly spikes | 37K |
 
 ## Usage
 
@@ -48,10 +48,12 @@ This dataset contains 10 years of global internet censorship measurements from 1
 from datasets import load_dataset
 
 # Load historical measurements
-ds = load_dataset("emperor-mew/global-censorship-index", data_files="ooni-historical.parquet")
+ds = load_dataset("emperor-mew/ooni-censorship-historical",
+                  data_files="data/ooni-historical.parquet")
 
 # Load labeled incidents (for ML training)
-incidents = load_dataset("emperor-mew/global-censorship-index", data_files="censorship-incidents.parquet")
+incidents = load_dataset("emperor-mew/ooni-censorship-historical",
+                         data_files="data/censorship-incidents.parquet")
 ```
 
 ## Schema
@@ -81,41 +83,55 @@ incidents = load_dataset("emperor-mew/global-censorship-index", data_files="cens
 
 ## Known Events Covered
 
-- 🇮🇷 Iran Mahsa Amini protests (2022)
-- 🇲🇲 Myanmar military coup (2021)
-- 🇧🇾 Belarus election shutdown (2020)
-- 🇷🇺 Russia Ukraine invasion blocks (2022+)
-- 🇰🇿 Kazakhstan January protests (2022)
-- 🇸🇩 Sudan military coup (2021)
-- 🇨🇺 Cuba July protests (2021)
-- 🇺🇬 Uganda election shutdown (2021)
-- And 17+ more...
+- Iran Mahsa Amini protests (2022)
+- Myanmar military coup (2021)
+- Belarus election shutdown (2020)
+- Russia Ukraine invasion blocks (2022+)
+- Kazakhstan January protests (2022)
+- Sudan military coup (2021)
+- Cuba July protests (2021)
+- Uganda election shutdown (2021)
+- And 17+ more
 
-## Model
+## Voidly Atlas ML Stack (2026-05-21)
 
-We provide a trained GradientBoosting classifier:
-- **F1 Score**: 99.8%
-- **ROC AUC**: 1.000
-- Available via API: `https://api.voidly.ai/hydra/v1/detect`
+This historical archive is the long-horizon training substrate for the
+**Voidly Atlas** ML stack. The production stack is documented in dedicated
+HuggingFace model cards under [`emperor-mew`](https://huggingface.co/emperor-mew):
+
+- **Classifier v3.3** ([`emperor-mew/voidly-classifier-v3.3`](https://huggingface.co/emperor-mew/voidly-classifier-v3.3)) — country-day censorship classifier, GradientBoosting, regime-similarity-weighted contagion features. Honest cross-country generalization: **leave-one-country-out median F1 0.87, mean F1 0.71**. The fitted `.pkl` + per-country thresholds ship in that repo.
+- **Multi-horizon forecast** ([`emperor-mew/voidly-forecast-v1-multi-horizon`](https://huggingface.co/emperor-mew/voidly-forecast-v1-multi-horizon)) — 1d/7d/30d XGBoost + isotonic, LOCO AUC 0.91 / 0.88 / 0.84.
+- **Unsupervised anomaly** ([`emperor-mew/voidly-anomaly-dbscan-v1`](https://huggingface.co/emperor-mew/voidly-anomaly-dbscan-v1)) — CenDTect-style DBSCAN second-opinion signal.
+- 12 more model cards — search `emperor-mew/voidly-` on the Hub.
+
+> **Note on the older "F1 99.8% / AUC 1.000" claim:** that figure was a
+> stratified-random-split number on a now-superseded v2 model. It does not
+> reflect cross-country generalization. The current honest metric is the
+> **LOCO** (leave-one-country-out) F1 reported above — random splits inflate
+> apparent accuracy because the model learns per-country base rates.
+
+For a clean held-out evaluation task, use the companion benchmark
+[`emperor-mew/voidly-bench-v1`](https://huggingface.co/datasets/emperor-mew/voidly-bench-v1).
 
 ## Citation
 
 ```bibtex
-@dataset{voidly_censorship_index_2026,
+@dataset{voidly_ooni_historical_2026,
   author = {Voidly Research},
-  title = {Global Censorship Index: 10 Years of Internet Measurement Data},
+  title = {Voidly OONI Censorship Historical: 10 Years of Internet Measurement Data},
   year = {2026},
   publisher = {Hugging Face},
-  url = {https://huggingface.co/datasets/emperor-mew/global-censorship-index}
+  url = {https://huggingface.co/datasets/emperor-mew/ooni-censorship-historical}
 }
 ```
 
 ## Links
 
-- 🌐 [Voidly Censorship Index](https://voidly.ai/censorship-index)
-- 📡 [Real-time API](https://api.voidly.ai/data/censorship-index.json)
-- 🤖 [MCP Server](https://www.npmjs.com/package/@voidly/mcp-server)
-- 📊 [OONI (source)](https://ooni.org)
+- [Voidly Censorship Index](https://voidly.ai/censorship-index)
+- [Real-time API](https://api.voidly.ai/data/censorship-index.json)
+- [MCP Server](https://www.npmjs.com/package/@voidly/mcp-server)
+- [Voidly Atlas ML model cards](https://huggingface.co/emperor-mew)
+- [OONI (source)](https://ooni.org)
 
 ## License