File size: 4,876 Bytes
8ff305b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b80c35e
8ff305b
b80c35e
 
8ff305b
 
 
 
 
 
2d4602a
 
 
 
 
 
 
 
 
 
 
 
 
 
296ac84
 
 
 
 
 
 
 
 
 
 
 
 
b80c35e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
dataset_info:
  features:
  - name: id
    dtype: string
  - name: table_type
    dtype: string
  - name: domain
    dtype: string
  - name: difficulty
    dtype: string
  - name: sample_index
    dtype: int32
  - name: query
    dtype: string
  - name: answer
    dtype: string
  - name: positive_image
    dtype: image
  - name: negative_image
    dtype: image
  - name: target_cell
    dtype: string
  - name: negative_type
    dtype: string
  splits:
  - name: train
    num_bytes: 430815279.0
    num_examples: 2000
  download_size: 427887702
  dataset_size: 430815279.0
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---
# Table retrieval training data

2,000 synthetic hard table retrieval pairs generated with GPT-6 Luna: 500 each of Vertical, Matrix, Hierarchy, and Nested.

Each train row contains a query, its answer, a positive image, a style-matched hard negative image, the table type and domain, target cell metadata (JSON string), and negative transformation type. Both image columns embed PNG bytes in Parquet shards.

Negative rules: Vertical swaps distinct values in one column; Matrix moves the queried entities onto the same row or column axis; Hierarchy swaps parent groups; Nested swaps outer groups. Rendering uses randomized Matplotlib, HTML/CSS, and LaTeX styles, preserved between each positive and negative.

`metadata/` preserves source table specifications, query templates and answers, original render styles, and hard-negative records. Paths inside these original records describe the source layout; use the embedded image columns when loading the dataset.

```python
from datasets import load_dataset
ds = load_dataset("NLPRETGOOD/table_train", split="train")
```

## Version v1.1.0: randomized Matrix entity positions

Matrix 엔티티의 행·열 위치를 무작위화했습니다. Gold tables independently shuffle rows and columns while preserving every entity-pair value. Hard negatives keep the same axis-swap rule and select their valid exchange partner randomly. All 500 Matrix pairs (1,000 PNGs) were rerendered with seed `20261109`; Matrix queries now use the visible table title. The remaining 1,500 training rows are unchanged.

See [CHANGELOG.md](CHANGELOG.md) and [versions/v1.1.0.json](versions/v1.1.0.json) for source image hashes, position counts, provenance and the baseline revision. Future releases use a new immutable `vX.Y.Z` tag and append their changes to the changelog.

Pin this version for reproducible training:

```python
from datasets import load_dataset
ds = load_dataset("NLPRETGOOD/table_train", revision="v1.1.0", split="train")
```

## v1.2.0 — 2026-11-09: Hierarchy/Nested position refresh

- Hierarchy 500쌍과 Nested 500쌍의 gold/HN을 재렌더링했습니다 (총 2,000 PNG, seed `20261109`).
- Hierarchy: parent 그룹 블록과 하위 행을 무작위로 재배치하고 다른 parent 중 swap 상대를 균등 확률로 선택합니다. 유한 표본의 위치별 개수는 정확히 같을 필요가 없습니다.
- Nested: target의 subtable·행·열 위치를 무작위 균등 배정합니다. 500개 target은 subtable별 125개, 행별 100개, 열별 166–167개입니다. 같은 target subtable 위치 안에서 swap 상대별 횟수 차이는 최대 1입니다.
- 기존 query, 정답, target identity, 스타일과 HN 관계 변경 규칙을 유지했습니다. Vertical/Matrix 1,000행은 `v1.1.0`와 완전히 같습니다.
- Rendered specs, HN metadata, position audits, provenance and verification reports are synchronized. Original authored specs remain unchanged.
- Parent revision: `296ac842567d45aefe3cda1e20bee02336dc8133`.

English: rerender 500 Hierarchy and 500 Nested gold/HN pairs to remove fixed display positions. Hierarchy uses seeded random parent/sibling permutations and a uniformly sampled other parent; Nested uses balanced randomized target positions and conditionally balanced swap partners. All query/answer/non-image fields and all Vertical/Matrix rows remain unchanged.

### Reproducibility and metadata

The default split remains `train` (2,000 rows, four table types × 500). Pin this release with:

```python
from datasets import load_dataset
ds = load_dataset("NLPRETGOOD/table_train", revision="v1.2.0", split="train")
```

`metadata/specs/` contains the unchanged authored specs, whose order is not the final rendered order. The exact rendered gold and HN specs are `metadata/rendered_assets/hard/gpt-6-luna-hard-train-500/<type>/randomized_specs.json` and `metadata/rendered_assets/hard/gpt-6-luna-hard-train-500/hn/<type>/randomized_specs.json`. Position audits and render provenance are alongside the gold specs. See [CHANGELOG.md](CHANGELOG.md), [release manifest](versions/v1.2.0.json), and [independent verification](verification/v1.2.0/independent_verification.json). Prior version tags are retained.