Add dataset card

README.md

@@ -3,6 +3,7 @@ language:
 - en
 license: cc-by-4.0
 task_categories:
+- conversational
 - text-generation
 tags:
 - anime
@@ -18,18 +19,6 @@ configs:
   data_files:
   - split: train
     path: data/train-*
-dataset_info:
-  features:
-  - name: user
-    dtype: string
-  - name: anime_girl
-    dtype: string
-  splits:
-  - name: train
-    num_bytes: 1080136
-    num_examples: 4000
-  download_size: 642851
-  dataset_size: 1080136
 ---
 
 # Anime Girl Chat Dataset (English)
@@ -40,20 +29,20 @@ character's `anime_girl` reply.
 
 ## Dataset details
 
-- **Rows:** 2000
+- **Rows:** 4000
 - **Columns:** `user` (string), `anime_girl` (string)
 - **Language:** English
 - **Format:** plain text only, no emoji or kaomoji, ASCII only
 - **Style layers:**
-  - **Sweet layer** (first 1000 rows): short, warm replies of 95-195 characters using `~` and small `*action*` beats.
-  - **PolyBuzz-style layer** (last 1000 rows): longer replies of 200-350 characters, each opening with a third-person narration block in asterisks (for example `*She looks up from the pot, steam curling around her cheeks*`) followed by warm deredere dialogue.
+  - **Sweet layer** (first half): short, warm replies of 90-230 characters using `~` and small `*action*` beats.
+  - **PolyBuzz-style layer** (second half): longer replies of 201-350 characters, each opening with a third-person narration block in asterisks (for example `*She looks up from the pot, steam curling around her cheeks*`) followed by warm deredere dialogue.
 
 Note for fine-tuning: the two layers are different voice intensities. For a single
 consistent persona, train on one layer or mix them deliberately.
 
 ## Splits
 
-- `train`: 2000 rows
+- `train`: 4000 rows
 
 ## Usage