Update README: rename script, remove balanced column, add citation

README.md

@@ -77,7 +77,7 @@ Smaller subsets are strict subsets of larger ones. Hours are measured by
 │   ├── s500hr_*.tar
 │   └── s1000hr_*.tar
 └── scripts/
-    └── tsv_to_metadata_shard.py   # TSV -> Kimi-Audio training JSONL
+    └── tsv_to_metadata.py   # TSV -> Kimi-Audio training JSONL
 ```
 
 Inside each tar, paths preserve the original tree:
@@ -176,13 +176,13 @@ df_50hr = subset(df, "50hr")   # equivalent to metadata_50hr.tsv
 
 ## Convert to Kimi-Audio training format
 
-The included script `tsv_to_metadata_shard.py` converts any metadata TSV
+The included script `tsv_to_metadata.py` converts any metadata TSV
 into the conversation-format JSONL used to replicate the
 [Kimi-Audio](https://github.com/MoonshotAI/Kimi-Audio) speech-to-speech
 training pipeline described in our paper.
 
 ```bash
-python scripts/tsv_to_metadata_shard.py metadata_100hr.tsv \
+python scripts/tsv_to_metadata.py metadata_100hr.tsv \
   -o metadata_shard.jsonl
 ```