Update README.md

README.md

@@ -133,3 +133,18 @@ configs:
   - split: train
     path: data/train-*
 ---
+
+# dataproc5/test-danbooru2025-tag-balanced-210k
+
+A test dataset of balanced 2100k rows from danbooru2025, aiming to maximize common knowledege per image under limited data or compute scope.
+
+```python
+balanced_df = select_balanced_images(dbr_df, tag_counts_monthly_df, tags_thres=1800, max_images=500_000, max_tag_proportion=0.08, random_seed=42)
+ub.saves(balanced_df, "hf://dataproc5/test-danbooru2025-tag-balanced-210k", private=False)
+```
+
+The `balanced-60k` subset is a moderate-sized dataset compared to [balanced-2k](https://huggingface.co/datasets/dataproc5/test-danbooru2025-tag-balanced-2k) or [balanced-10k](https://huggingface.co/datasets/dataproc5/test-danbooru2025-tag-balanced-10k) or [balanced-60k](https://huggingface.co/datasets/dataproc5/test-danbooru2025-tag-balanced-60k), and is targeted for:
+
+- medium-sized model finetunes
+- quick model arch experiments
+- classifier / aesthetic predictor evaluations