victormiller
commited on
Commit
•
497eac9
1
Parent(s):
64d8614
Update README.md
Browse files
README.md
CHANGED
@@ -52,15 +52,16 @@ Curated datasets are typically structured and consistently formatted. TxT360 fil
|
|
52 |
| Europarl | 6.1 GB | 1.96B | - |
|
53 |
| StackExchange | 45 GB | 8.37B | Q4 2023 |
|
54 |
|
|
|
55 |
|
56 |
## CommonCrawl Data Filtering
|
57 |
-
|
58 |
|
59 |
## Curated Source Filtering
|
60 |
-
Each data source was filtered individually with respect to the underlying data. Full details and discussion on how each source
|
61 |
|
62 |
## Global Deduplication
|
63 |
-
After the web and curated sources were filtered,
|
64 |
|
65 |
# Citation
|
66 |
|
|
|
52 |
| Europarl | 6.1 GB | 1.96B | - |
|
53 |
| StackExchange | 45 GB | 8.37B | Q4 2023 |
|
54 |
|
55 |
+
The [TxT360](https://huggingface.co/spaces/LLM360/TxT360) blog post provides all the details behind how we approached and implemented the following features:
|
56 |
|
57 |
## CommonCrawl Data Filtering
|
58 |
+
Complete discussion on how 99 Common Crawl snapshots were filtered and comparison to previous filtering techinques (e.g. Dolma, DataTrove, RedPajamaV2).
|
59 |
|
60 |
## Curated Source Filtering
|
61 |
+
Each data source was filtered individually with respect to the underlying data. Full details and discussion on how each source was filter are covered.
|
62 |
|
63 |
## Global Deduplication
|
64 |
+
After the web and curated sources were filtered, all sources globally deduplicated to create TxT360. The tips and tricks behind the deduplication process are included.
|
65 |
|
66 |
# Citation
|
67 |
|