Add the re-recap dataset card

README.md

@@ -0,0 +1,84 @@
+---
+license: cc-by-4.0
+language:
+- en
+pretty_name: re-recap
+size_categories:
+- 1M<n<10M
+task_categories:
+- image-to-text
+- text-to-image
+tags:
+- captions
+- image-text
+- recap-datacomp
+---
+
+# re-recap
+
+re-recap is a table of live image URLs, each with two captions: the Recap-DataComp caption that came with the URL, and a longer caption written from the image by [Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B).
+
+This repository currently holds the dataset card only. The parquet shards are not published yet.
+
+## What a row contains
+
+Each published row has exactly three fields.
+
+| Field | What it is |
+| --- | --- |
+| `url` | A web image URL that returned a decodable image when this set was built. |
+| `recap_caption` | The Recap-DataComp recaption for that URL (`re_caption` in [UCSC-VLAA/Recap-DataComp-1B](https://huggingface.co/datasets/UCSC-VLAA/Recap-DataComp-1B)). |
+| `phi9_caption` | A new caption written by Qwen3.8-27B from the image at `url`. |
+
+The short original web alt-text, CLIP scores, GPT-4V ratings, hashes, and image bytes are not included.
+
+## How the rows were chosen
+
+The source is the first 10 million rows of a live-URL keep of Recap-DataComp-1B. Recap points at DataComp-1B web images and does not store pixels. Many of those URLs are dead, so this set keeps a URL only when a full download returned an image that decoded. Rows that failed to download or decode are dropped, which is why the published count will be smaller than 10 million. The exact count will be written here when the shards are uploaded.
+
+`recap_caption` is copied from Recap. It is not rewritten.
+
+## How the Phi-9 caption was written
+
+For each URL that still returned an image:
+
+1. The image was downloaded.
+2. If the longer side was above 1280 pixels, the whole picture was scaled down so that side became 1280. Smaller images were left at their original size. The full frame was kept.
+3. That image was shown to [Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B). The Recap caption was not shown with it.
+4. The model was asked for a plain visual description of 150 to 200 words, in two short paragraphs, including any words visible in the picture. Captions longer than 200 words were cut back to that length.
+
+The downloaded images were used only to write the caption. They are not stored in this dataset.
+
+## License
+
+The URL list and `recap_caption` are a subset of Recap-DataComp-1B, which is released under [CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/). Recap itself redistributes URL–text metadata from DataComp-1B, also under CC-BY-4.0. This dataset is therefore released under CC-BY-4.0.
+
+That license covers the table in this repository: the URLs and both caption strings. It does not cover the images. Each image stays under whatever copyright applies at its URL. Downloading an image is your own responsibility and is subject to the site that hosts it. DataComp states the same limit: the URL–text index is CC-BY-4.0, and the individual images remain under their own copyrights.
+
+Attribution for the inherited fields:
+
+- Li et al., *What If We Recaption Billions of Web Images with LLaMA-3?*, ICML 2025. [arXiv:2406.08478](https://arxiv.org/abs/2406.08478). Dataset: [UCSC-VLAA/Recap-DataComp-1B](https://huggingface.co/datasets/UCSC-VLAA/Recap-DataComp-1B).
+- Gadre et al., *DataComp: In search of the next generation of multimodal datasets*, NeurIPS 2023. Dataset: [mlfoundations/datacomp_1b](https://huggingface.co/datasets/mlfoundations/datacomp_1b).
+
+`phi9_caption` was written with Qwen3.8-27B, whose weights are Apache-2.0. Those captions are included in this CC-BY-4.0 release.
+
+## Limitations
+
+URLs go stale. A link that worked when the caption was written may 404 later.
+
+The images are ordinary web photographs and graphics, with the same noise, duplicates, and safety residue as DataComp and Recap. This release does not add a new safety filter.
+
+`phi9_caption` can still misread small text, counts, or identities. It is a visual description, not a verified fact sheet. `recap_caption` was written by a different, smaller model (LLaVA-1.5 with LLaMA-3-8B, about 50 tokens, from the Recap paper) and will disagree with the Phi-9 caption on some rows.
+
+## Citation
+
+```bibtex
+@misc{phi9_re_recap_2026,
+  title        = {re-recap},
+  author       = {{Phi-9}},
+  year         = {2026},
+  howpublished = {\url{https://huggingface.co/datasets/phi-9/re-recap}}
+}
+```
+
+If you use `recap_caption` or the URLs, also cite Li et al. (2024/2025) and Gadre et al. (2023).