Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
|
@@ -1,6 +1,6 @@
|
|
| 1 |
# Document Haystack Dataset
|
| 2 |
|
| 3 |
-
This repository contains the dataset for the paper
|
| 4 |
|
| 5 |
---
|
| 6 |
|
|
@@ -18,6 +18,14 @@ It expands on the *Needle in a Haystack* concept by embedding *needles* — shor
|
|
| 18 |
|
| 19 |
---
|
| 20 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 21 |
## ✅ Key Features
|
| 22 |
|
| 23 |
- **Document Lengths:** 5, 10, 25, 50, 75, 100, 150, 200 pages
|
|
@@ -33,16 +41,12 @@ It expands on the *Needle in a Haystack* concept by embedding *needles* — shor
|
|
| 33 |
|
| 34 |
---
|
| 35 |
|
| 36 |
-
##
|
| 37 |
|
| 38 |
-
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
| Text+Image Needles # Documents | 25 | 25 | 25 | 25 | 25 | 25 | 25 | 25 | 200 |
|
| 43 |
-
| Text+Image Needles # Questions | 125 | 250 | 625 | 625 | 625 | 625 | 625 | 625 | 4125 |
|
| 44 |
-
| **Total Documents** | **50** | **50** | **50** | **50** | **50** | **50** | **50** | **50** | **400** |
|
| 45 |
-
| **Total Questions** | **250** | **500** | **1250** | **1250** | **1250** | **1250** | **1250** | **1250** | **8250** |
|
| 46 |
|
| 47 |
---
|
| 48 |
|
|
@@ -58,6 +62,19 @@ It expands on the *Needle in a Haystack* concept by embedding *needles* — shor
|
|
| 58 |
|
| 59 |
---
|
| 60 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 61 |
## 📁 Dataset Structure
|
| 62 |
|
| 63 |
Below is an example of the dataset’s folder layout:
|
|
@@ -151,24 +168,6 @@ The secret office supply is a "pencil".,2,8,gray,white,0.339,0.931,times-bold,90
|
|
| 151 |
|
| 152 |
---
|
| 153 |
|
| 154 |
-
## 🎯 Benchmark Design
|
| 155 |
-
|
| 156 |
-
- **Key-value pairs:** Needles are in the format: “The secret *KEY* is *VALUE*.” (e.g., “The secret sport is *basketball*.”). Cover diverse categories (sports, animals, currencies, fruits, instruments, etc.) — see Table 3 in paper for full list.
|
| 157 |
-
- **Tasks:** Each needle has an associated retrieval question:
|
| 158 |
-
_“What is the secret KEY in the document?”_
|
| 159 |
-
- **Objective scoring:** The VLM’s answer is checked for the correct `VALUE` (or acceptable aliases for text+image needles).
|
| 160 |
-
|
| 161 |
-
---
|
| 162 |
-
|
| 163 |
-
## 🔬 Use Cases
|
| 164 |
-
|
| 165 |
-
- Stress-test multimodal VLMs for long-context understanding
|
| 166 |
-
- Compare retrieval from parsed pdf text vs images vs original pdfs
|
| 167 |
-
- Explore text vs image vs mixed retrieval challenges
|
| 168 |
-
- Measure performance drop with increasing context length
|
| 169 |
-
|
| 170 |
-
---
|
| 171 |
-
|
| 172 |
## 📏 Reference Evaluation
|
| 173 |
|
| 174 |
Use with the [Document Haystack Benchmark Code](https://github.com/amazon-science/document-haystack) for:
|
|
|
|
| 1 |
# Document Haystack Dataset
|
| 2 |
|
| 3 |
+
This repository contains the dataset for the paper “Document Haystack: A Long Context Multimodal Image/Document Understanding Vision LLM Benchmark”.
|
| 4 |
|
| 5 |
---
|
| 6 |
|
|
|
|
| 18 |
|
| 19 |
---
|
| 20 |
|
| 21 |
+
## 🎯 Benchmark Design
|
| 22 |
+
|
| 23 |
+
- **Key-value pairs:** Each needle follows the pattern “The secret *KEY* is *VALUE*.” where `VALUE` appears as either text or an image. For example: “The secret sport is *basketball*.”. The keys span diverse categories including sports, animals, currencies, fruits, musical instruments, and more (see Table 3 in the paper for the complete category list).
|
| 24 |
+
- **Tasks:** Each needle has an associated retrieval question: “What is the secret *KEY* in the document?”
|
| 25 |
+
- **Objective scoring:** The VLM’s answer is checked for the correct `VALUE` (or acceptable aliases for text+image needles).
|
| 26 |
+
|
| 27 |
+
---
|
| 28 |
+
|
| 29 |
## ✅ Key Features
|
| 30 |
|
| 31 |
- **Document Lengths:** 5, 10, 25, 50, 75, 100, 150, 200 pages
|
|
|
|
| 41 |
|
| 42 |
---
|
| 43 |
|
| 44 |
+
## 🔬 Use Cases
|
| 45 |
|
| 46 |
+
- Stress-test multimodal VLMs for long-context understanding
|
| 47 |
+
- Compare retrieval from parsed pdf text vs images vs original pdfs
|
| 48 |
+
- Explore text vs image vs mixed retrieval challenges
|
| 49 |
+
- Measure performance drop with increasing context length
|
|
|
|
|
|
|
|
|
|
|
|
|
| 50 |
|
| 51 |
---
|
| 52 |
|
|
|
|
| 62 |
|
| 63 |
---
|
| 64 |
|
| 65 |
+
## 📊 Document Haystack Characteristics
|
| 66 |
+
|
| 67 |
+
| # Pages | 5 | 10 | 25 | 50 | 75 | 100 | 150 | 200 | Total |
|
| 68 |
+
|----------------|-----|-----|------|------|------|-------|-------|-------|-------|
|
| 69 |
+
| Text Needles # Documents | 25 | 25 | 25 | 25 | 25 | 25 | 25 | 25 | 200 |
|
| 70 |
+
| Text Needles # Questions | 125 | 250 | 625 | 625 | 625 | 625 | 625 | 625 | 4125 |
|
| 71 |
+
| Text+Image Needles # Documents | 25 | 25 | 25 | 25 | 25 | 25 | 25 | 25 | 200 |
|
| 72 |
+
| Text+Image Needles # Questions | 125 | 250 | 625 | 625 | 625 | 625 | 625 | 625 | 4125 |
|
| 73 |
+
| **Total Documents** | **50** | **50** | **50** | **50** | **50** | **50** | **50** | **50** | **400** |
|
| 74 |
+
| **Total Questions** | **250** | **500** | **1250** | **1250** | **1250** | **1250** | **1250** | **1250** | **8250** |
|
| 75 |
+
|
| 76 |
+
---
|
| 77 |
+
|
| 78 |
## 📁 Dataset Structure
|
| 79 |
|
| 80 |
Below is an example of the dataset’s folder layout:
|
|
|
|
| 168 |
|
| 169 |
---
|
| 170 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 171 |
## 📏 Reference Evaluation
|
| 172 |
|
| 173 |
Use with the [Document Haystack Benchmark Code](https://github.com/amazon-science/document-haystack) for:
|