Upload README.md with huggingface_hub

README.md

@@ -1,41 +1,14 @@
 ---
-dataset_info:
-  features:
-  - name: utt_id
-    dtype: string
-  - name: speaker
-    dtype: string
-  - name: ref
-    dtype: string
-  - name: hyp
-    dtype: string
-  - name: n_ref_words
-    dtype: int64
-  - name: n_errors
-    dtype: int64
-  - name: wer_pct
-    dtype: float64
-  - name: n_substitutions
-    dtype: int64
-  - name: n_deletions
-    dtype: int64
-  - name: n_insertions
-    dtype: int64
-  - name: substitution_pairs
-    dtype: string
-  - name: error_classes
-    dtype: string
-  - name: rank
-    dtype: int64
-  splits:
-  - name: train
-    num_bytes: 70775
-    num_examples: 100
-  download_size: 50678
-  dataset_size: 70775
-configs:
-- config_name: default
-  data_files:
-  - split: train
-    path: data/train-*
+license: cc-by-4.0
+task_categories: [automatic-speech-recognition]
+language: [en]
+pretty_name: Diffusion-ASR hardest 100 test-clean clips
 ---
+# The 100 test-clean clips with the most errors (diffusion-ASR, 4.60% system)
+
+The 100 clips (of 2,620) carrying the most word errors - 547 errors, 22.4% of the
+system's total - ranked worst-first (`rank` column). Same columns as
+[the full dataset](https://huggingface.co/datasets/Tachyeon/diffusion-asr-testclean-decodes).
+Composition (measured): long sentences (median 37 words) dense with rare words and names
+(12.8% out-of-dictionary vs 9.5% corpus-wide); 34 distinct narrators, so not a narrator
+artifact. Join `utt_id` against LibriSpeech test-clean (CC BY 4.0) for audio.