---
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
- recaption
- vision-language
configs:
- config_name: short
  data_files:
  - split: preview
    path: short/preview.parquet
  - split: train
    path: short/train-*.parquet
  default: true
- config_name: long
  data_files:
  - split: preview
    path: long/preview.parquet
  - split: train
    path: long/train-*.parquet
dataset_info:
- config_name: short
  features:
  - name: url
    dtype: string
  - name: recap_caption
    dtype: string
  - name: phi9_caption
    dtype: string
  splits:
  - name: preview
    num_bytes: 65077
    num_examples: 100
  - name: train
    num_bytes: 4295495465
    num_examples: 8369267
  download_size: 4295560542
  dataset_size: 4295560542
- config_name: long
  features:
  - name: url
    dtype: string
  - name: recap_caption
    dtype: string
  - name: phi9_caption
    dtype: string
  splits:
  - name: preview
    num_bytes: 114441
    num_examples: 100
  - name: train
    num_bytes: 1319058800
    num_examples: 1348204
  download_size: 1319173241
  dataset_size: 1319173241
---

# re-recap

**re-recap** is an open vision-language dataset pairing live web image URLs with dual captions: the original baseline recaption from [Recap-DataComp-1B](https://huggingface.co/datasets/UCSC-VLAA/Recap-DataComp-1B) and an expanded, highly detailed visual description generated directly from image pixels using [Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B).

The dataset contains **9,717,471** verified image-text pairs partitioned into two distinct subsets based on caption length and prompt structure.

## Overview and Dataset Subsets

To serve different modeling needs—ranging from concise text-to-image conditioning to exhaustive scene understanding—the dataset is organized into two configurations:

| Subset | Examples | Typical Length | Description |
| :--- | :---: | :---: | :--- |
| **`short`** *(default)* | 8,369,267 | 150–200 words | Structured, two-paragraph scene descriptions focused on primary subjects, environment layout, lighting, and legible text, constrained to a maximum of 200 words. |
| **`long`** | 1,348,204 | ~400 words | Comprehensive, open-ended visual descriptions capturing fine-grained texture, background elements, spatial relationships, and extensive detail without word limits. |

Each subset provides a lightweight **`preview`** split containing 100 rows for immediate visual inspection in the Hugging Face Dataset Viewer, alongside the complete **`train`** split containing the full parquet collection.

## Data Fields

Each row in the dataset contains exactly three string columns:

- `url` (`string`): The live web URL from which the image was successfully retrieved and validated during pipeline processing.
- `recap_caption` (`string`): The unmodified baseline recaption distributed in Recap-DataComp-1B, originally produced by an 8-billion parameter vision-language model (LLaVA-1.5 with LLaMA-3-8B) with an average length of approximately 50 tokens.
- `phi9_caption` (`string`): The new, objective description synthesized by Qwen3.8-27B after downloading and inspecting the image at `url`.

Superficial web alt-text, quality heuristics, CLIP similarity scores, internal hashes, and raw image pixels are omitted to ensure a compact, clean tabular release.

## Curation and Processing Pipeline

The source pool comprises the first 10 million rows of a live-URL verification pass over Recap-DataComp-1B. Because web-crawled datasets suffer from link decay over time, every URL was actively tested prior to caption generation:

1. **Network Ingestion and Verification**: Each URL was fetched over HTTP. If a link timed out, returned an HTTP error, or produced non-image content such as HTML error pages, the row was dropped.
2. **Image Preprocessing and Framing**: Downloaded image bytes were decoded and converted to standard RGB. If either dimension exceeded 1280 pixels, the entire image was proportionally scaled down until its maximum side measured 1280 pixels. Smaller images retained their original resolution. No cropping was applied, preserving the full compositional frame and camera perspective.
3. **Model Generation**: The preprocessed image was passed to Qwen3.8-27B. The original Recap caption was strictly excluded from the prompt context, ensuring that the model relied purely on observable visual evidence rather than prior synthetic summaries. The prompt instructed the model to remain factual, describe subjects and spatial geometry objectively, and transcribe visible text verbatim within quotation marks.
4. **Length Conditioning**: The first 1.35 million captions were generated without length limits, producing the `long` subset. The subsequent 8.37 million captions were generated under an explicit 150-to-200 word guidance target and a 200-word truncation ceiling, forming the `short` subset.

After caption generation was completed, temporary image buffers were discarded from memory. No image binaries are hosted or redistributed in this repository.

## Licensing and Usage Rights

The metadata table in this repository—consisting of the URL mappings, the inherited Recap captions, and the generated descriptions—is distributed under the [Creative Commons Attribution 4.0 International License (CC-BY-4.0)](https://creativecommons.org/licenses/by/4.0/).

Please note the following boundaries regarding data rights:

- **Image Copyright**: The underlying images are not owned, hosted, or licensed by Phi-9. Individual images remain under the copyright and terms of service of their respective origin websites. Any party downloading images via the provided URLs assumes all responsibility regarding usage and compliance with source copyright terms.
- **Model Outputs**: The generated captions were synthesized using open Apache-2.0 weights from Qwen3.8-27B and are contributed to the public domain under CC-BY-4.0.

### Attribution

When utilizing the inherited URLs and baseline captions, please credit the original dataset curators:

- 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).

## Limitations and Quality Considerations

- **URL Longevity**: Although every link decoded successfully during pipeline execution, external web servers change over time, and some URLs may subsequently become unavailable.
- **Web Domain Noise**: The source images stem from Common Crawl web data, inheriting the duplicate content, compression artifacts, and varied distributions typical of large-scale web harvests.
- **Model Hallucinations**: While Qwen3.8-27B demonstrates strong visual fidelity, automated descriptions can occasionally misread small typography, conflate subtle background objects, or misjudge exact counts in crowded scenes.

## Citation

```bibtex
@misc{phi9_re_recap_2026,
  title        = {re-recap: 9.7M Live Web Image-Caption Pairs with Detailed Qwen3.8 Synthesized Descriptions},
  author       = {{Phi-9}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/phi-9/re-recap}}
}
```