Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
|
@@ -779,19 +779,6 @@ original `PaDaS-Lab/webfaq` shards; `kept` = rows remaining after cleaning.
|
|
| 779 |
|
| 780 |
## Cleaning
|
| 781 |
|
| 782 |
-
Each row goes through four lightweight, mostly substring-based passes:
|
| 783 |
-
a substring blocklist (cross-language SEO doorway markers like KAYAK /
|
| 784 |
-
momondo / Tripadvisor / TrickTactoe, plus per-language patterns for
|
| 785 |
-
zh and ja); a yes/no echo + query-template filter (zh/ja-specific);
|
| 786 |
-
near-duplicate dedup via MinHash LSH (64 perms, 16 bands, char-5gram
|
| 787 |
-
shingles, Jaccard ≥ 0.85 per bucket); and encoding repair (HTML
|
| 788 |
-
entities, JSON-unicode-escape leftovers, and literal `\r\n\t` fixed
|
| 789 |
-
in place; rows still containing `\ufffd` after fix are dropped). The
|
| 790 |
-
pipeline is conservative — it targets templated SEO noise and known
|
| 791 |
-
encoding glitches rather than trying to score semantic relevance.
|
| 792 |
-
|
| 793 |
-
## Cleaning
|
| 794 |
-
|
| 795 |
Filters were developed iteratively: sample random rows per language, have
|
| 796 |
an LLM inspect them to identify recurring failure patterns (templated SEO
|
| 797 |
doorways, casino doorway pages, mismatched product-listing Q–Ds, encoding
|
|
|
|
| 779 |
|
| 780 |
## Cleaning
|
| 781 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 782 |
Filters were developed iteratively: sample random rows per language, have
|
| 783 |
an LLM inspect them to identify recurring failure patterns (templated SEO
|
| 784 |
doorways, casino doorway pages, mismatched product-listing Q–Ds, encoding
|