Publish data/CC-MAIN-2026-08/00000.parquet

README.md

@@ -32,7 +32,7 @@ configs:
 
 Open Index is a large-scale web text dataset built from [Common Crawl](https://commoncrawl.org). Every page goes through a pipeline that extracts the main content from raw HTML, converts it to clean Markdown using [trafilatura](https://github.com/adbar/trafilatura), and packages the result into Parquet files with full WARC metadata preserved.
 
-The dataset currently includes crawl **CC-MAIN-2026-08**. We plan to add more snapshots over time.
+The dataset currently includes crawl **CC-MAIN-2026-08** with **232,408 documents across 11 shards**. We plan to add more snapshots over time.
 
 Open Index is released under the **Open Data Commons Attribution License (ODC-By) v1.0**, the same license used by Common Crawl.
 
@@ -166,18 +166,18 @@ Empty conversions (pages where trafilatura could not extract meaningful content)
 
 ### Compression Ratios
 
-The pipeline achieves aggressive size reduction at every stage. Numbers below are measured from shard `00000` of CC-MAIN-2026-08 (~21,000 pages), and projected to the full crawl.
+Numbers below are measured from 11 shards of CC-MAIN-2026-08.
 
-| Stage | Per shard | Projected (100,000 shards) | Reduction |
-|---|---|---|---|
-| Raw WARC (.warc.gz, downloaded) | ~830 MB | ~83 TB | — |
-| HTML extracted (uncompressed) | 2.7 GB | ~270 TB | — |
-| Packed markdown WARC (.md.warc.gz) | 36.3 MB | ~3.6 TB | **-97.8%** vs raw WARC |
-| Final Parquet (ZSTD level 19) | 27.9 MB | ~2.8 TB | **-23%** vs packed WARC |
+| Stage | Per shard (avg) | Reduction |
+|---|---|---|
+| Raw WARC (.warc.gz, downloaded) | ~830 MB | — |
+| HTML extracted (uncompressed) | 2.9 GB | — |
+| Packed markdown WARC (.md.warc.gz) | 79.9 MB | **-97.3%%** vs HTML |
+| Final Parquet (Zstd level 19) | 28.7 MB | **-64.1%%** vs packed WARC |
 
-The big win happens in the HTML → Markdown conversion step: trafilatura strips all HTML tags, inline scripts, stylesheets, navigation elements, footers, and ads, keeping only the main content. This cuts the uncompressed text from 2.7 GB to 79 MB per shard — a **97.2% reduction** — before any file-level compression is applied.
+The big win is the HTML → Markdown step: trafilatura strips all tags, scripts, styles, navigation, and ads, keeping only the main content. This cuts uncompressed HTML from 2.9 GB to 79.9 MB per shard — a **97.3%% reduction** — before any file-level compression.
 
-Parquet with Zstd level 19 then adds another 23% on top of the already-compressed WARC gzip. End to end, one raw WARC shard of ~830 MB becomes a 27.9 MB Parquet file containing ~21,000 clean markdown documents.
+End to end: one raw WARC shard of ~830 MB becomes a **28.7 MB Parquet file** containing ~21128 clean markdown documents.
 
 ### Personal and Sensitive Information
 

stats.csv

@@ -0,0 +1,12 @@
+crawl_id,file_idx,rows,html_bytes,md_bytes,parquet_bytes
+CC-MAIN-2026-08,0,21184,2934018801,83014521,29226909
+CC-MAIN-2026-08,1,21265,2961325911,80724752,29324414
+CC-MAIN-2026-08,2,21141,2920084882,77766365,28624679
+CC-MAIN-2026-08,3,21086,2959834291,80424489,28852889
+CC-MAIN-2026-08,4,21124,2984315198,79569270,28811454
+CC-MAIN-2026-08,5,21118,2940457892,79695450,28359001
+CC-MAIN-2026-08,6,20999,2926359938,79996642,28536135
+CC-MAIN-2026-08,7,21161,2926551572,79361922,28897356
+CC-MAIN-2026-08,8,21155,2969697387,81276956,28586435
+CC-MAIN-2026-08,9,21131,2910850749,77004269,28060998
+CC-MAIN-2026-08,10,21044,2933010991,80544092,28836695