---
pretty_name: SciFormaBench-2K Full-Reference Backup
license: other
task_categories:
- text-to-image
language:
- en
size_categories:
- 1K<n<10K
tags:
- scientific-diagrams
- benchmark
- evaluation
- methodology-figures
configs:
- config_name: simple
  data_files:
  - split: test
    path: data/simple.parquet
- config_name: medium
  data_files:
  - split: test
    path: data/medium.parquet
- config_name: hard
  data_files:
  - split: test
    path: data/hard.parquet
---

<h1 align="center">SciFormaBench-2K: Full-Reference Benchmark Backup</h1>

<p align="center">
  <a href="https://arxiv.org/abs/2607.18091"><img alt="arXiv" src="https://img.shields.io/badge/arXiv-2607.18091-b31b1b?logo=arxiv&logoColor=white" height="22" /></a>
  &nbsp;
  <a href="https://github.com/microsoft/SciForma"><img alt="Code" src="https://img.shields.io/badge/GitHub-Code-181717?logo=github&logoColor=white" height="22" /></a>
  &nbsp;
  <a href="https://microsoft.github.io/SciForma/"><img alt="Project Page" src="https://img.shields.io/badge/Project-Page-2ea44f?logo=githubpages&logoColor=white" height="22" /></a>
  &nbsp;
  <a href="https://huggingface.co/microsoft/SciForma-9B"><img alt="Models" src="https://img.shields.io/badge/%F0%9F%A4%97-Models-yellow" height="22" /></a>
</p>

This repository is the complete-reference backup of **SciFormaBench-2K**, the
official evaluation benchmark for
[SciForma](https://github.com/microsoft/SciForma). It contains all 2,000
reference images, prompts, target resolutions, provenance fields, and
Component/Arrow/Text structural rubrics.

## Features

- 🖼️ **Complete reference set.** All 2,000 ground-truth scientific
  methodology diagrams are embedded in the Parquet files.
- 🧩 **Official benchmark splits.** Simple/Medium/Hard contain
  500/900/600 samples with the original benchmark indices and ordering.
- 🎯 **Axis-specific evaluation.** Each sample includes independently
  verified Component, Arrow, and Text inventories.
- 🔗 **Source-linked records.** Each row retains its arXiv ID, year, source
  image path, prompt, and target dimensions.
- ✅ **Complete text rubrics.** `rubric_text` is populated from the official
  `stage1_abstraction.text_labels`; 1,873 rows contain non-empty text-label
  inventories.

## Usage

```python
import json
from datasets import load_dataset

dataset = load_dataset(
    "LoYuXrqw/SciFormaBench-2K",
    "hard",                    # simple / medium / hard
    split="test",
    token=True,
)

sample = dataset[0]
sample["gt_image"].show()

components = json.loads(sample["rubric_components"])
arrows = json.loads(sample["rubric_arrows"])
text_labels = json.loads(sample["rubric_text"])
print(sample["index"], len(components), len(arrows), len(text_labels))
```

Set `HF_TOKEN` or log in with `hf auth login` before loading this repository.

## Evaluation

Use the official [SciForma evaluation pipeline](https://github.com/microsoft/SciForma/tree/main/eval):

```bash
python eval/eval_benchmark.py \
  --gen_dir ./results/my_model \
  --output_dir ./eval_results/my_model \
  --hf_benchmark LoYuXrqw/SciFormaBench-2K \
  --levels hard \
  --num_retest 2 \
  --deployment_name gpt-5.4
```

Use `--levels simple medium hard` for the complete benchmark. The paper uses
GPT-5.4 as the judge; GPT-4o is the public alternative. The official protocol
uses axis-specific calls with `split_dims=True` and averages two evaluation
rounds with `num_retest=2`.

## License

This full-reference repository is maintained for internal preservation and
evaluation. It does not provide a new redistribution license for the embedded
figures. Each image remains governed by its original source-paper license.