Create README.md
Browse files
README.md
CHANGED
@@ -0,0 +1,19 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: mit
|
3 |
+
---
|
4 |
+
|
5 |
+
We release the raw passages, embeddings, and index of MassiveDS.
|
6 |
+
|
7 |
+
Website: https://retrievalscaling.github.io
|
8 |
+
|
9 |
+
We release two versions of MassiveDS:
|
10 |
+
1. [MassiveDS-1.4T](https://huggingface.co/datasets/rulins/MassiveDS-1.4T), which contains 1.4T tokens in the datastore.
|
11 |
+
2. [MassiveDS-140B](https://huggingface.co/datasets/rulins/MassiveDS-140B), which is a subsampled version containing 140B tokens in the datastore.
|
12 |
+
|
13 |
+
File structure:
|
14 |
+
* `raw_data`: plain data in JSONL files.
|
15 |
+
* `passages`: chunked raw passages with passage IDs. Each passage is chunked to have no more than 256 words.
|
16 |
+
* `embeddings`: embeddings of the passages encoded with Contriever-MSMACRO.
|
17 |
+
* `index`: flat index built with embeddings.
|
18 |
+
|
19 |
+
Note: due to large data volume, we are still uploading the data for MassiveDS-1.4T. MassiveDS-140B is ready to go. Code support to run with MassiveDS is in https://github.com/RulinShao/retrieval-scaling.
|