Remove unsourced rule grouping table
Browse files
README.md
CHANGED
|
@@ -78,16 +78,21 @@ Each Common Crawl dump is processed independently:
|
|
| 78 |
7. **Duplicate removal:** remove duplicate followers using the generated blacklist.
|
| 79 |
8. **Text sanitation:** repair residual HTML entities and mojibake, remove invisible characters, and reject corrupted or abnormally sized documents.
|
| 80 |
|
| 81 |
-
|
| 82 |
|
| 83 |
-
|
| 84 |
-
|:---|:---|
|
| 85 |
-
| Gopher Repetition | Duplicate lines and paragraphs, dominant n-grams, repeated-character ratios |
|
| 86 |
-
| Gopher Quality | Document length, mean word length, symbol ratios, stop-word and alphabetic-word coverage |
|
| 87 |
-
| C4 | Punctuation, boilerplate, JavaScript fragments, `lorem ipsum`, and policy text |
|
| 88 |
-
| FineWeb Quality | Short lines, line-ending punctuation, repeated-line characters, and newline density |
|
| 89 |
|
| 90 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 91 |
|
| 92 |
## 📁 Dataset Structure
|
| 93 |
|
|
|
|
| 78 |
7. **Duplicate removal:** remove duplicate followers using the generated blacklist.
|
| 79 |
8. **Text sanitation:** repair residual HTML entities and mojibake, remove invisible characters, and reject corrupted or abnormally sized documents.
|
| 80 |
|
| 81 |
+
Deduplication is performed within each Common Crawl dump rather than across the full collection.
|
| 82 |
|
| 83 |
+
## 📈 Comparison with FineWeb
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 84 |
|
| 85 |
+
As a reproducibility reference, we compare URL sets on `CC-MAIN-2025-26`. After MinHash deduplication, the Ultra-FineWeb-L1 output shares 175,281,630 URLs with FineWeb. This is a pipeline comparison for one validation dump, not a quality ranking or a requirement that the two datasets contain the same documents.
|
| 86 |
+
|
| 87 |
+
| Metric | Before MinHash | After MinHash |
|
| 88 |
+
|:---|---:|---:|
|
| 89 |
+
| Ultra-FineWeb-L1 unique URLs | 258,071,345 | 194,402,835 |
|
| 90 |
+
| FineWeb unique URLs | 193,438,138 | 193,438,138 |
|
| 91 |
+
| URL intersection | 182,421,184 | 175,281,630 |
|
| 92 |
+
| Recall against FineWeb | 94.30% | **90.61%** |
|
| 93 |
+
| Precision | 70.69% | **90.16%** |
|
| 94 |
+
|
| 95 |
+
After MinHash deduplication, precision increases while recall decreases. This reflects a trade-off in URL coverage and deduplication; it is not a direct measurement of overall text quality. These figures apply only to the `CC-MAIN-2025-26` validation run and are not general claims about all released dumps.
|
| 96 |
|
| 97 |
## 📁 Dataset Structure
|
| 98 |
|