Add Gemma 3 4B IT recaptions for the first 500,000 images

gemma-3-4b-it-recaptions/README.md

@@ -0,0 +1,25 @@
+# Gemma 3 4B IT recaptions
+
+One new caption per image for the first 500,000 pairs of this dataset, in subset order (`index` 0–499,999). The images and original captions stay in the tar shards. Join on `key`.
+
+These captions were written by [google/gemma-3-4b-it](https://huggingface.co/google/gemma-3-4b-it) with vLLM 0.33.0 (bfloat16, greedy decoding, images capped at 896 px on the long side, pan-and-scan off, at most 220 new tokens). Two A6000 GPUs, about 4.3 images/s each.
+
+| Column | Content |
+| --- | --- |
+| `key` | example key in the tar |
+| `source_shard` | tar file name |
+| `recap` | the new caption, one paragraph |
+| `words` | word count of `recap` |
+| `index` | position in the 500,000-pair subset (0–499,999) |
+
+`webdataset_shard_289.parquet` has the first 9,088 images of that shard. Later shards in the 1,000,000-pair subset are not included.
+
+```python
+from datasets import load_dataset
+
+recaps = load_dataset(
+    "parquet",
+    data_files="hf://datasets/IntelligenceLab/BLIP3o-Recap/gemma-3-4b-it-recaptions/*.parquet",
+    split="train",
+)
+```