File size: 6,061 Bytes
95762a0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b8e28e6
95762a0
 
 
 
b8e28e6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
95762a0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
---
license: other
license_name: mixed-per-source-see-below
license_link: https://huggingface.co/datasets/bhaskar1225/verifydocbench#licensing-by-source
task_categories:
- token-classification
- document-question-answering
language:
- en
- de
- es
- fr
- zh
tags:
- document-extraction
- calibration
- selective-prediction
- conformal-prediction
- grounding
- key-information-extraction
- trustworthy-ml
pretty_name: VerifyDocBench
size_categories:
- 10K<n<100K
---

# VerifyDocBench

Genuine per-field trust records for document-extraction calibration, risk-controlled
selective prediction, and grounding research. Companion resource to:

- **VerifyDocBench** (benchmark paper) — first released benchmark scoring per-field
  calibration, risk-controlled selective prediction, and grounding-conditioned trust
  for document extraction.
- **"Valid Per-Field Selective Risk Control for Document Extraction"** (method paper) —
  the validity ladder / conformal risk-control procedures this data was captured to
  evaluate.

Code, loaders, scoring harness, and both papers' reproducibility artifacts:
<https://github.com/bhaskargurram-ai/verifydoc> (Apache-2.0).

## What's in this repository

Nine genuine per-field capture files, **36,265 records total**, from four extractor
families over four corpora:

| file | extractor | fields | docs | source corpus |
|---|---|---|---|---|
| `cord_claude-sonnet-5.json` | claude-sonnet-5 (text-layer) | 13,859 | 800 | CORD (receipts) |
| `funsd_claude-sonnet-5.json` | claude-sonnet-5 (text-layer) | 1,999 | 175 | FUNSD (forms) |
| `xfund_de_claude-sonnet-5.json` | claude-sonnet-5 (text-layer) | 523 | 42 | XFUND German |
| `xfund_es_claude-sonnet-5.json` | claude-sonnet-5 (text-layer) | 632 | 43 | XFUND Spanish |
| `xfund_fr_claude-sonnet-5.json` | claude-sonnet-5 (text-layer) | 772 | 44 | XFUND French |
| `xfund_zh_claude-sonnet-5.json` | claude-sonnet-5 (text-layer) | 461 | 27 | XFUND Chinese |
| `cord_claude-haiku-4-5.json` | claude-haiku-4-5 (text-layer) | 5,341 | 400 | CORD (frozen-config confirmation) |
| `cord_qwen2.5-14b.json` | Qwen2.5-14B (vLLM, open-weights, text-layer) | 6,168 | 398 | CORD |
| `cord_gpt-4o.json` | **gpt-4o (genuinely vision-based)** | 6,510 | 399 | CORD |

**Only `cord_gpt-4o.json` is genuinely vision-based** — the model was sent the
rendered page image (base64-encoded) alongside the OCR text layer as assistive
context. Every other file is **text-layer prompted**: the model reads only the
document's OCR text, never pixels, despite "VLM" being the natural shorthand for a
closed frontier API model — we correct that framing here rather than let it stand
for three of the four extractor families.

Each record is one field prediction:

```json
{
  "doc_id": "cord-train-00000",
  "path": "menu.nm",
  "value": "Nasi Campur Bali",
  "verbalized": 0.95,
  "consistency": 1.0,
  "grounded": true,
  "support": 0.83,
  "correct": 1,
  "entailment": 0.91
}
```

`verbalized` = the model's self-reported confidence; `consistency` = k-sample
self-consistency agreement; `grounded`/`support` = whether/how strongly the value was
located in the document's text layer (ambiguity-penalized); `entailment` = NLI
cross-encoder score for "does the grounded span entail this value"; `correct` = scored
against gold (see the paper for exact-match/numeric/semantic scoring rules per field
type).

**Not included, and not claimed as released:** DocILE and SROIE — no genuine capture
exists for either in this project (DocILE download is pending a valid access token;
SROIE has loader code but was never run for a real capture). Source document *images*
are not re-hosted for any corpus (see licensing below) — only our own derived per-field
predictions and scoring.

## Licensing by source

**Our own contributions** — the per-field correctness labels, grounding/support scores,
schemas, and this derived record format — are released under **Apache-2.0**, matching
the parent repository's license.

This does **not** relicense the underlying source document collections, each of which
keeps its own license:

| source | license | notes |
|---|---|---|
| CORD | **CC BY 4.0** | Permissive; our Apache-2.0 annotations are consistent with it. |
| FUNSD | **Non-commercial, research/educational use only** (custom terms) | Not a CC license. Use of `funsd_claude-sonnet-5.json` must stay within these terms. |
| XFUND (de/es/fr/zh) | **CC BY-NC-SA 4.0** | Non-commercial, share-alike; uniform across languages. |

We redistribute only our own added annotations (values, confidence/grounding/
correctness signals) and reference the original dataset downloads for source
documents/images — restrictively licensed source images are never re-hosted here.
If you redistribute derivatives of the FUNSD or XFUND slices, you must comply with
their non-commercial terms.

## Reproducibility

Captured via the harness at `scripts/apivlm_perfield_rich.py` (seed-pinned, k=3
self-consistency sampling, doc-level splits, regression-gated). Every number in both
companion papers traces to these files or to the scoring harness's deterministic
transforms of them — see `paper/HANDOFF.md` and `paper/generated/real-runs/` in the
GitHub repository for the full campaign log.

## Human-gold audit

A subset of these records was independently re-judged by three blind human annotators
(Fleiss' κ=0.83 on 600 CORD/FUNSD items, κ=0.939 on a further 400 XFUND items) to bound
automatic-label noise. See the method paper's confirmation section and
`benchmark/card.md` in the GitHub repository for the full audit.

## Citation

```bibtex
@misc{verifydocbench2026,
  title  = {VerifyDocBench: Measuring Per-Field Calibration, Selective Risk, and Grounding for Document Extraction},
  author = {Gurram, Bhaskar},
  year   = {2026},
  url    = {https://github.com/bhaskargurram-ai/verifydoc}
}
```

## Maintenance

Maintained by the repository owner via GitHub issues:
<https://github.com/bhaskargurram-ai/verifydoc/issues>. No formal versioning/erratum
policy beyond the issue tracker as of this release.