Update README.md

README.md

@@ -5,24 +5,26 @@ size_categories:
 - n<1K
 task_categories:
 - image-classification
+- visual-question-answering
+- visual-document-retrieval
 task_ids: []
 pretty_name: document-haystack-10pages
 tags:
 - fiftyone
 - image
-- image-classification
-dataset_summary: '
+dataset_summary: >
 
 
 
 
-  This is a [FiftyOne](https://github.com/voxel51/fiftyone) dataset with 250 samples.
+  This is a [FiftyOne](https://github.com/voxel51/fiftyone) dataset with 250
+  samples.
 
 
   ## Installation
 
 
-  If you haven''t already, install FiftyOne:
+  If you haven't already, install FiftyOne:
 
 
   ```bash
@@ -44,7 +46,7 @@ dataset_summary: '
 
   # Load the dataset
 
-  # Note: other available arguments include ''max_samples'', etc
+  # Note: other available arguments include 'max_samples', etc
 
   dataset = load_from_hub("harpreetsahota/document-haystack-10pages")
 
@@ -54,16 +56,12 @@ dataset_summary: '
   session = fo.launch_app(dataset)
 
   ```
-
-  '
+license: cc-by-nc-4.0
 ---
 
 # Dataset Card for document-haystack-10pages
 
-<!-- Provide a quick summary of the dataset. -->
-
-
-
+[!image/png](document_haystacks_fo.gif)
 
 
 This is a [FiftyOne](https://github.com/voxel51/fiftyone) dataset with 250 samples.
@@ -95,130 +93,211 @@ session = fo.launch_app(dataset)
 
 ### Dataset Description
 
-<!-- Provide a longer summary of what this dataset is. -->
+Document Haystack is a comprehensive benchmark designed to evaluate the performance of Vision Language Models (VLMs) on long, visually complex documents. This FiftyOne dataset contains the 5-page subset, which serves as an entry point for testing retrieval capabilities on shorter documents.
 
+The benchmark expands on the "Needle in a Haystack" concept by embedding needles — short key-value statements in pure text or as multimodal text+image snippets — within real-world documents. These needles test whether models can locate specific information hidden within complex documents with textual, visual, or mixed content.
 
+**Key Features:**
+- 25 real-world base documents (annual reports, financial filings, etc.)
+- 5 pages per document variant
+- 5 needles per document (strategically placed across pages)
+- Two needle types: text-only and text+image
+- 250 total samples (125 per needle type)
+- 250 retrieval questions
 
-- **Curated by:** [More Information Needed]
-- **Funded by [optional]:** [More Information Needed]
-- **Shared by [optional]:** [More Information Needed]
-- **Language(s) (NLP):** en
-- **License:** [More Information Needed]
+- **Curated by:** Amazon AGI (Goeric Huybrechts, Srikanth Ronanki, Sai Muralidhar Jayanthi, Jack Fitzgerald, Srinivasan Veeravanallur)
+- **Language:** English
+- **License:** CC-BY-NC-4.0
 
-### Dataset Sources [optional]
+### Dataset Sources
 
-<!-- Provide the basic links for the dataset. -->
+- **Original Repository:** https://github.com/amazon-science/document-haystack
+- **Original Dataset:** https://huggingface.co/datasets/AmazonScience/document-haystack
+- **Paper:** [Document Haystack: A Long Context Multimodal Image/Document Understanding Vision LLM Benchmark](https://arxiv.org/abs/2507.15882)
 
-- **Repository:** [More Information Needed]
-- **Paper [optional]:** [More Information Needed]
-- **Demo [optional]:** [More Information Needed]
+## Dataset Structure
 
-## Uses
+### FiftyOne Schema
 
-<!-- Address questions around how the dataset is intended to be used. -->
+Each sample in this FiftyOne dataset represents a single page image with the following fields:
 
-### Direct Use
+#### Sample-Level Fields
 
-<!-- This section describes suitable use cases for the dataset. -->
+| Field | Type | Description |
+|-------|------|-------------|
+| `filepath` | string | Path to the page image (JPG, 200 DPI) |
+| `document_name` | Classification | Name of the source document (e.g., "AIG", "AmericanAirlines") |
+| `page_number` | int | Page number within the document (1-5) |
+| `needle_type` | Classification | Type of needles: "text" or "text_image" |
 
-[More Information Needed]
+#### Needle Information Fields
 
-### Out-of-Scope Use
+These fields contain lists of information about needles on each page:
 
-<!-- This section addresses misuse, malicious use, and uses that the dataset will not work well for. -->
+| Field | Type | Description |
+|-------|------|-------------|
+| `needle_texts` | list[string] | Full needle statements (e.g., "The secret currency is a \"euro\".") |
+| `needle_keys` | Classifications | Extracted keys (e.g., "currency", "sport") |
+| `needle_answers` | Classifications | Extracted answers (e.g., "euro", "basketball") |
+| `needle_questions` | list[string] | Questions for retrieving each needle |
+| `needle_font_sizes` | list[int] | Font sizes of needles |
+| `needle_text_colors` | list[string] | Text colors |
+| `needle_bg_colors` | list[string] | Background colors |
+| `needle_fonts` | list[string] | Font families |
+| `needle_scales` | list[int] | Scale values (for text+image needles) |
+| `needle_locations` | Keypoints | Spatial locations of needles with (x, y) coordinates in [0, 1] |
 
-[More Information Needed]
+### Needle Categories
 
-## Dataset Structure
-
-<!-- This section provides a description of the dataset fields, and additional information about the dataset structure such as criteria used to create the splits, relationships between data points, etc. -->
+Needles span diverse categories including:
+- Sports
+- Animals
+- Currencies
+- Fruits
+- Musical instruments
+- Office supplies
+- Flowers
+- Landmarks
+- And more...
 
-[More Information Needed]
+### Needle Format
 
-## Dataset Creation
+Each needle follows the pattern: **"The secret KEY is VALUE."**
 
-### Curation Rationale
+- **Text needles:** Both KEY and VALUE are text (e.g., "The secret sport is basketball.")
+- **Text+Image needles:** VALUE is shown as an image (e.g., "The secret sport is [image of basketball]")
 
-<!-- Motivation for the creation of this dataset. -->
+Questions follow the pattern: **"What is the secret KEY in the document?"**
 
-[More Information Needed]
+## Uses
 
-### Source Data
+### Direct Use
 
-<!-- This section describes the source data (e.g. news text and headlines, social media posts, translated sentences, ...). -->
+This dataset is designed for:
 
-#### Data Collection and Processing
+1. **Evaluating VLM retrieval capabilities** - Test how well models can locate specific information within documents
+2. **Benchmarking long-context understanding** - Even at 5 pages, this tests models' ability to process extended visual content
+3. **Comparing text vs. multimodal retrieval** - Direct comparison between text-only and text+image needle performance
+4. **Visual dataset exploration** - Use FiftyOne's visualization tools to understand needle placement patterns
+5. **Model development** - Train and validate models for document understanding tasks
 
-<!-- This section describes the data collection and processing process such as data selection criteria, filtering and normalization methods, tools and libraries used, etc. -->
+### Out-of-Scope Use
 
-[More Information Needed]
+- This dataset is for **research and evaluation purposes only** (CC-BY-NC-4.0 license)
+- Not intended for commercial use without proper licensing
+- Not suitable for training models to extract sensitive information from documents
+- The 5-page subset is not representative of truly long-context scenarios (use 50-200 page subsets for that)
 
-#### Who are the source data producers?
+## Dataset Creation
 
-<!-- This section describes the people or systems who originally created the data. It should also include self-reported demographic or identity information for the source data creators if this information is available. -->
+### Curation Rationale
 
-[More Information Needed]
+The Document Haystack benchmark was created to address the lack of suitable benchmarks for evaluating VLMs on long, visually complex documents. While many benchmarks focus on perception tasks, processing long documents with both text and visual elements remains under-explored.
 
-### Annotations [optional]
+The 5-page subset serves as:
+- An accessible entry point for initial model testing
+- A faster iteration benchmark for development
+- A baseline for comparing against longer document performance
 
-<!-- If the dataset contains annotations which are not part of the initial data collection, use this section to describe them. -->
+### Source Data
 
-#### Annotation process
+#### Data Collection and Processing
 
-<!-- This section describes the annotation process such as annotation tools used in the process, the amount of data annotated, annotation guidelines provided to the annotators, interannotator statistics, annotation validation, etc. -->
+- **Base documents:** Real-world documents including annual reports, financial filings, and corporate documents from 25 different organizations
+- **Page extraction:** Documents converted to 200 DPI page images
+- **Needle insertion:** Key-value pairs strategically placed across pages with controlled randomization
+  - Needles placed in non-overlapping page ranges
+  - Same locations used for both text and text+image variants
+  - Various visual properties (fonts, colors, sizes) to test robustness
+- **Text extraction:** OCR/parsing for text-only variants
 
-[More Information Needed]
+#### Who are the source data producers?
 
-#### Who are the annotators?
+The original documents are publicly available corporate documents (annual reports, financial statements, etc.). The benchmark itself was created by researchers at Amazon AGI.
 
-<!-- This section describes the people or systems who created the annotations. -->
+### Annotations
 
-[More Information Needed]
+#### Annotation Process
 
-#### Personal and Sensitive Information
+Annotations include:
+- **Needle placement metadata:** Precise coordinates, font properties, colors
+- **Ground truth answers:** Extracted key-value pairs for each needle
+- **Retrieval questions:** Automatically generated questions following the template "What is the secret KEY in the document?"
 
-<!-- State whether the dataset contains data that might be considered personal, sensitive, or private (e.g., data that reveals addresses, uniquely identifiable names or aliases, racial or ethnic origins, sexual orientations, religious beliefs, political opinions, financial or health data, etc.). If efforts were made to anonymize the data, describe the anonymization process. -->
+The placement is automated but controlled to ensure:
+- Even distribution across document pages
+- Non-overlapping placement
+- Visibility and retrievability
 
-[More Information Needed]
+#### Who are the Annotators?
 
-## Bias, Risks, and Limitations
+Annotations were automatically generated as part of the benchmark creation process by the Amazon AGI research team.
 
-<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+### Personal and Sensitive Information
 
-[More Information Needed]
+The dataset consists of publicly available corporate documents (annual reports, financial filings). No personal or sensitive information beyond what is publicly available in these corporate documents is present.
 
-### Recommendations
+Needles are synthetic key-value pairs and do not contain real sensitive information.
 
-<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+## Bias, Risks, and Limitations
 
-Users should be made aware of the risks, biases and limitations of the dataset. More information needed for further recommendations.
+**Limitations:**
+- **Document diversity:** Limited to 25 base documents from corporate/financial domain
+- **English-only:** All documents and needles are in English
+- **5-page constraint:** Not representative of truly long documents (50-200 pages)
+- **Synthetic task:** Needle retrieval is a proxy task for real document understanding
+- **Visual styling:** Needles use specific visual properties that may not represent all real-world scenarios
 
-## Citation [optional]
+**Biases:**
+- Corporate document focus may not generalize to other document types
+- Needle categories reflect common Western concepts
+- Single language limits cross-lingual evaluation
 
-<!-- If there is a paper or blog post introducing the dataset, the APA and Bibtex information for that should go in this section. -->
+**Risks:**
+- Models optimized for this benchmark may overfit to the needle retrieval pattern
+- Performance on this task may not correlate with general document understanding
 
-**BibTeX:**
+### Recommendations
 
-[More Information Needed]
+Users should:
+- Use this dataset alongside other document understanding benchmarks
+- Test on multiple page lengths (5, 10, 25, 50+ pages) for comprehensive evaluation
+- Consider domain shift when applying models to non-corporate documents
+- Validate that good performance translates to real-world document tasks
+- Be aware this tests retrieval, not deeper comprehension or reasoning
 
-**APA:**
+## Citation
 
-[More Information Needed]
+If you use this dataset, please cite the original Document Haystack paper:
 
-## Glossary [optional]
+**BibTeX:**
 
-<!-- If relevant, include terms and calculations in this section that can help readers understand the dataset or dataset card. -->
+```bibtex
+@article{huybrechts2025document,
+  title={Document Haystack: A Long Context Multimodal Image/Document Understanding Vision LLM Benchmark},
+  author={Huybrechts, Goeric and Ronanki, Srikanth and Jayanthi, Sai Muralidhar and Fitzgerald, Jack and Veeravanallur, Srinivasan},
+  journal={arXiv preprint arXiv:2507.15882},
+  year={2025}
+}
+```
 
-[More Information Needed]
+**APA:**
 
-## More Information [optional]
+Huybrechts, G., Ronanki, S., Jayanthi, S. M., Fitzgerald, J., & Veeravanallur, S. (2025). Document Haystack: A Long Context Multimodal Image/Document Understanding Vision LLM Benchmark. arXiv preprint arXiv:2507.15882.
 
-[More Information Needed]
+## More Information
 
-## Dataset Card Authors [optional]
+### FiftyOne Integration
 
-[More Information Needed]
+This dataset leverages FiftyOne's capabilities for:
+- Visual exploration of needle locations via Keypoints
+- Filtering and querying by document properties
+- Classification-based analysis of keys and answers
+- Integration with FiftyOne Brain for embeddings and similarity
 
-## Dataset Card Contact
+### Related Datasets
 
-[More Information Needed]
+The full Document Haystack benchmark includes variants with:
+- 5, 10, 25, 50, 75, 100, 150, and 200 pages
+- 400 document variants total
+- 8,250 retrieval questions