Update README.md

README.md

@@ -1,25 +1,44 @@
----
-license: cc
-configs:
-- config_name: default
-  data_files:
-  - split: train
-    path: data/train-*
-dataset_info:
-  features:
-  - name: prompt
-    dtype: string
-  - name: audio_code
-    dtype:
-      array2_d:
-        shape:
-        - null
-        - 8
-        dtype: int16
-  splits:
-  - name: train
-    num_bytes: 357436897
-    num_examples: 7916
-  download_size: 180406553
-  dataset_size: 357436897
----
+---
+license: cc
+configs:
+- config_name: default
+  data_files:
+  - split: train
+    path: data/train-*
+dataset_info:
+  features:
+  - name: prompt
+    dtype: string
+  - name: audio_code
+    dtype:
+      array2_d:
+        shape:
+        - null
+        - 8
+        dtype: int16
+  splits:
+  - name: train
+    num_bytes: 357436897
+    num_examples: 7916
+  download_size: 180406553
+  dataset_size: 357436897
+task_categories:
+- text-to-audio
+- audio-classification
+language:
+- en
+tags:
+- music
+- art
+pretty_name: fmaEncodec
+size_categories:
+- 1K<n<10K
+---
+# Free Music Archive (FMA) small encodec version
+
+This dataset is built from Free Music Archive (FMA) small version ( https://huggingface.co/datasets/benjamin-paine/free-music-archive-small ). 
+We combine the metadata in the original dataset into prompt, and use Encodec-24kHZ (https://huggingface.co/facebook/encodec_24khz) 
+to encode the 30s audio into 2d array of shape (8,2249) or (8,2251) or the like.
+
+- "prompt": a description of the music, including title, author, genres, etc.
+- "audio_code": 2d array of shape [T, 8], T≈2249, indices are integers from 0 to 1023, encoded by `encodec_24khz`