File size: 5,502 Bytes
47b3bd7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
---
pretty_name: AURA OCR Benchmark — Real5-OmniDocBench Illumination (2026-11-20)
task_categories:
  - image-to-text
tags:
  - ocr
  - document-layout-analysis
  - omnidocbench
  - benchmark
size_categories:
  - 1K<n<10K
---

# AURA OCR Benchmark — Real5-OmniDocBench Illumination

This dataset holds the results of a two-stage document OCR pipeline run on the
**illumination** subset of [Real5-OmniDocBench](https://huggingface.co/datasets/PaddlePaddle/Real5-OmniDocBench):

- **Stage 1, layout:** the AURA page endpoint finds regions on each page.
- **Stage 2, OCR:** AURA OCR reads each region, then the regions are assembled into Markdown.

The run was made on 2026-11-20. It uses **worst-case concurrency settings**, so the speed figures are an upper bound on cost.

## Results

1,354 of 1,355 pages were scored. Scores use OmniDocBench metrics, evaluated against the official `OmniDocBench.json` ground truth.

| Metric | Value |
|---|---|
| **Overall** ↑ | **83.66** |
| Text edit distance ↓ | 0.1788 |
| Formula CDM ↑ | 76.66 |
| Table TEDS ↑ | 92.19 |
| Reading-order edit distance ↓ | 0.2273 |

Overall = ((1 − text edit) × 100 + table TEDS + formula CDM) / 3.

### Speed and projected effort

| Stage | Seconds per file | Mandays for 2.192M files |
|---|---|---|
| Layout (AURA `/v1/ocr/page`) | 3.84 | 97.4 |
| OCR (AURA OCR, per region) | 3.42 | 86.7 |
| **Total** | **7.26** | **184.1** |

- **Seconds per file:** wall-clock time divided by files, at the concurrency settings listed below. One file is one page in this subset.
- **Mandays:** 2,192,000 files × seconds per file ÷ 86,400 (24-hour days).
- **Combined throughput:** 0.14 pages per second.
- **The two stages ran one after the other.** Running them in a pipeline (overlapped) could bring the total close to the slower stage alone, if the API limits allow both at once.
- **Client machine:** a laptop with a 12th Gen Intel Core i5-12500H. All inference runs on the remote AURA endpoint, and peak client RAM was under 1 GB.

## Run configuration

| Setting | Value |
|---|---|
| Layout endpoint | `POST /v1/ocr/page?format=paddlex&ocr=false` (boxes only, with per-box confidence scores) |
| Layout post-processing | GLM-OCR NMS and containment merge; duplicate boxes dropped |
| Layout concurrency | 3 pages in flight (the endpoint's per-key limit) |
| OCR model | `aura-ocr`, one request per region |
| OCR prompts | `Text Recognition:` for text, `Table Recognition:` for tables, `Formula Recognition:` for formulas. Figures are not OCR'd; they are embedded as image crops. |
| OCR concurrency | 8 files in parallel, 1 region at a time per file (`AURA_REGION_WORKERS=1`, worst case) |
| Rate limit | 500 requests per minute, enforced client-side |
| Max output tokens | 4096 per request |

### Layout statistics

- **Regions:** 29,302 detected across 1,355 pages, about 21.6 per page.
- **By category:** text 24,154; figure 1,742; other 1,533; formula 1,303; table 570.
- **Pages by document type:**
  - jiaocaineedrop 214
  - docstructbench 170
  - yanbaopptmerge 159
  - newspaper 151
  - notes 116
  - PPT 108
  - yanbaor2 107
  - book 72
  - exam 70
  - eastmoney 53
  - color textbook 42
  - magazine 42
  - scihub 39
  - jiaocai 12

## Repository layout

```
layout_results_AURA/                    # Stage 1 output (about 44 GB)
  summary_metrics.json                  # layout timing summary
  <doc_stem>/
    metadata.txt                        # per-file timing
    page_001/
      page.png                          # the page as sent to the endpoint
      layout.json                       # regions: label, category, score, bbox, reading order, crop file
      layout_vis.png                    # boxes drawn on the page
      crop_<n>_<category>_<x1>_<y1>_<x2>_<y2>.png
vllm_benchmark/                         # Stage 2 output
  VLLM_Report_51186.pdf                 # benchmark report
  vllm_benchmark_summary.json           # metrics for the report
  aura-ocr_layout_enabled_formula/
    omnidoc_benchmark_summary.{json,txt}   # accuracy summary
    errors.txt                          # files that failed
    <doc_stem>/
      markdown.md                       # final page Markdown
      metadata.txt                      # per-page latency and token details
      crops/                            # figure crops referenced by markdown.md (only on pages with figures)
aura_illum_layout.log                   # Stage 1 console log
aura_illum_ocr.log                      # Stage 2 console log
```

Each `layout.json` region looks like this:

```json
{"id": 1, "label": "paragraph_title", "cls_id": 17, "order": 0, "category": "text",
 "score": 0.8782, "bbox": [1586, 302, 2248, 438], "crop_filename": "crop_001_text_1586_302_2248_438.png"}
```

## Known issues

- **One page was not scored.** `color_textbook_…人教新起点英语5A电子课本_page_004` contains only illustrations with blacked-out labels. The layout endpoint found 0 regions, so the OCR stage logged it in `errors.txt` instead of saving an empty page.
- **Transient errors during OCR.** A few HTTP 429 (rate limit) responses and dropped connections happened; all succeeded on retry. One region returned no text after 3 retries and was saved as blank.
- **The layout run was resumed once.** The layout timing (3.84 s per file) comes from the resumed run, which covered 1,164 files; the first 191 files were skipped. That run also includes a few slow files caused by a short DNS outage, which add about 0.04 s per file.