Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,21 @@
|
|
| 1 |
-
---
|
| 2 |
-
license: cc
|
| 3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: cc
|
| 3 |
+
task_categories:
|
| 4 |
+
- text2text-generation
|
| 5 |
+
language:
|
| 6 |
+
- ja
|
| 7 |
+
tags:
|
| 8 |
+
- furigana
|
| 9 |
+
- education
|
| 10 |
+
pretty_name: 青空文庫振り仮名注釈付き音声コーパス(FLFL Train)
|
| 11 |
+
size_categories:
|
| 12 |
+
- 1M<n<10M
|
| 13 |
+
---
|
| 14 |
+
|
| 15 |
+
A more aggressively cleaned up version of [Calvin-Xu/Furigana-Aozora-Speech](https://huggingface.co/datasets/Calvin-Xu/Furigana-Aozora-Speech/blob/main/README.md). Training data of [Calvin-Xu/FLFL](https://huggingface.co/Calvin-Xu/FLFL/upload/main).
|
| 16 |
+
|
| 17 |
+
The Whisper-generated transcriptions in the original dataset contains many errors. Some sanity-checking is implemented to filter on reading of common kanji and eliminate wildly inaccurate entries.
|
| 18 |
+
|
| 19 |
+
Additionally as recognized by the original dataset, the consonant voicing (連濁) of many words, especially numbers / number + counters, is incorrect. Such words' readings were fixed in the training data (i.e., 十匹: じゅうひき -> じっぴき).
|
| 20 |
+
|
| 21 |
+
[Calvin-Xu/FLFL](https://huggingface.co/Calvin-Xu/FLFL/upload/main) is finetuned from [stockmark/gpt-neox-japanese-1.4b](https://huggingface.co/stockmark/gpt-neox-japanese-1.4b) and trained for slightly over one epoch on these data.
|