Update README.md

README.md

@@ -1,30 +1,45 @@
----
-license: odc-by
-configs:
-- config_name: default
-  data_files:
-  - split: train
-    path: data/train-*
-dataset_info:
-  features:
-  - name: text_1
-    dtype: string
-  - name: lang_1
-    dtype: string
-  - name: embedding_1
-    list: float32
-    length: 1024
-  - name: text_2
-    dtype: string
-  - name: lang_2
-    dtype: string
-  - name: embedding_2
-    list: float32
-    length: 1024
-  splits:
-  - name: train
-    num_bytes: 91746488145.0
-    num_examples: 10729806
-  download_size: 103633594260
-  dataset_size: 91746488145.0
----
+---
+license: odc-by
+configs:
+- config_name: default
+  data_files:
+  - split: train
+    path: data/train-*
+dataset_info:
+  features:
+  - name: text_1
+    dtype: string
+  - name: lang_1
+    dtype: string
+  - name: embedding_1
+    list: float32
+    length: 1024
+  - name: text_2
+    dtype: string
+  - name: lang_2
+    dtype: string
+  - name: embedding_2
+    list: float32
+    length: 1024
+  splits:
+  - name: train
+    num_bytes: 91746488145.0
+    num_examples: 10729806
+  download_size: 103633594260
+  dataset_size: 91746488145.0
+---
+
+[Original dataset](https://github.com/facebookresearch/fairseq/blob/nllb/examples/nllb/data/README.md)
+
+[SONAR's author message](https://github.com/facebookresearch/SONAR/issues/75#issuecomment-3221061689)
+
+What happened to the original dataset?
+We preprocess the public data and NLLB-Seed, with a note:
+1. `aau` dataset, except the `orm` language because it is not supported by the SONAR model
+2. `hornmt` dataset, except the `orm` and `aar` languages because these are not supported by the SONAR model
+3. `mburisano` dataset, except `nde` and `ven` languages because these are not supported by the SONAR model
+4. `tico` dataset, except the `orm` language because it is not supported by the SONAR model
+5. `til` dataset because the total size is about ~80.6M pairs, which translates to about ~615 GB
+
+What are the use cases?
+1. Training a model with embeddings as input. For example, training a Sparse Autoencoder. This saves computation because we do not need to load the encoder during training. Also, we do not need to cache the encoder's output on-the-fly