README.md

---
license: cc-by-4.0
task_categories:
- visual-question-answering
- image-to-text
tags:
- chart-qa
- vision-language
- synthetic
- reliability
- calibration
- self-consistency
pretty_name: RENDEQ
size_categories:
- 1K<n<10K
---

# RENDEQ

RENDEQ is a generator and dataset of **render-equivalence sets**: scientific
figures redrawn from the same underlying data under many cosmetically
different but semantically identical styles. Because the answer is computed
directly from the data by a deterministic program, every rendering of an
instance shares an exact, programmatically known answer — including exact
ground truth for the intermediate values a model must read off the figure,
not just the final answer.

This dataset accompanies the paper *"When Does Consensus Mean Correctness?
Measuring the Agreement–Accuracy Coupling with Semantics-Preserving
Re-Rendering."* It is used to measure whether a vision-language model's
agreement across re-renderings of a chart predicts whether its answer is
correct — a question that is normally impossible to test directly, since
ordinary image perturbations only preserve meaning by assumption and no
exact answer key exists to localize errors. RENDEQ removes both obstacles.

## What's in it

Three **independently generated instantiations** (`rendeq_v4`, `rendeq_v5`,
`rendeq_v6` — generator seeds 0, 1, 2), each with 350 instances across 7 plot
families and `K=8` renderings per instance (2,800 images per instantiation;
8,400 images total). Every headline result in the paper is pooled across all
three instantiations, never computed on a single one.

```
rendeq_v4/
  manifest.jsonl     # one JSON object per instance
  build_meta.json     # generator config for this instantiation (seed, style space, families)
  images/
    bar-extremum-0bd9daaebd17a04d__r0.png   # rendering 0 of 8
    bar-extremum-0bd9daaebd17a04d__r1.png
    ...
rendeq_v5/  (same structure, seed 1)
rendeq_v6/  (same structure, seed 2)
```

### Manifest schema

Each line of `manifest.jsonl` is one instance:

```json
{
  "instance_id": "bar-extremum-0bd9daaebd17a04d",
  "family": "bar",
  "question": {
    "qtype": "extremum",
    "text": "Which category has the highest value?",
    "answer_type": "categorical",
    "params": {"which": "max"}
  },
  "answer": {
    "value": "South",
    "answer_type": "categorical",
    "reads": {"South": 93.8, "Q4": 82.5, "Q3": 5.3, "Site D": 92.8, "Site C": 8.2}
  },
  "data_fingerprint": "0bd9daaebd17a04d",
  "renderings": [
    {"index": 0, "style": {"library": "plotly", "palette": "colorblind", ...}, "image": "bar-extremum-0bd9daaebd17a04d__r0.png"},
    ...
  ]
}
```

- **`answer.value`** — the exact answer, computed from the underlying data by
  a deterministic program, not by any model.
- **`answer.reads`** — the exact intermediate values the answer program
  consumed (e.g. every bar's height, not just which one is tallest). Used in
  the paper to separate perception failures (a misread value) from reasoning
  failures (every read correct, wrong final answer).
- **`renderings[i].style`** — the full style configuration for that rendering
  (plotting library, palette, theme, gridlines, aspect ratio, marker, font
  scale, tick density, and several family-specific factors). Every rendering
  of an instance shares the same `answer`, by construction — only `style`
  varies.
- **`data_fingerprint`** — identifies the underlying data `D`; all 8
  renderings of an instance share the same fingerprint.

### Plot families and axes

| Axis | Levels |
|---|---|
| Plot family | bar, grouped bar, stacked bar, line, scatter, pie, log-axis |
| Question type | value read, comparison, extremum, trend sign, arithmetic |
| Style Θ | library, palette, theme, gridlines, aspect ratio, marker, font, ticks, legend, resolution (plus family-specific factors: bar orientation/sort/labels, stack order, pie start angle/explode, line style, axis zero) |
| Difficulty | number of series, number of points, precision, near-ties, overlap |

Style groups are individually togglable, so a single nuisance factor (e.g.
plotting library) can be varied in isolation while holding everything else
fixed — this is what makes the per-factor attribution in the paper causal
with respect to style, not just correlational.

## Scope

This release is the **in-style synthetic split** used for every result in
the paper. The generator also supports an out-of-style split and a
real-data anchor built by re-rendering tables from a public benchmark;
neither was run for the paper (external validity to real charts is
untested — see the paper's Limitations), and neither is included here.

## Loading

```python
import json
from pathlib import Path
from huggingface_hub import snapshot_download

repo_dir = Path(snapshot_download(repo_id="rasulkhanbayov/rendeq", repo_type="dataset"))

manifest = [json.loads(l) for l in open(repo_dir / "rendeq_v4" / "manifest.jsonl")]
instance = manifest[0]
image_path = repo_dir / "rendeq_v4" / "images" / instance["renderings"][0]["image"]
```

## Generator

The generator that produced this data (`rendeq_generator.py`) is released
separately under the MIT license, alongside the full inference/scoring/
analysis pipeline used to produce every number in the paper. Regenerating
a fresh instantiation with a new seed, or building the out-of-style split,
only requires that code — this repository is the generator's *output*, not
the generator itself.

## License

CC-BY-4.0. All data is synthetic: category and series labels (e.g. "South",
"Q4", "Site D") are generator-assigned placeholders, not real-world entities,
and no personally identifying or offensive content is possible by
construction.

## Citation

The paper is currently under review; a citation will be added here once a
venue and citable version are available.