---
configs:
  - config_name: benchmark_432
    data_dir: "hf_viewer"
    drop_labels: true
---

# ChartBench-E Benchmark 432

This repository contains the frozen 432-sample benchmark build used in the paper:

> ChartBench-E: a benchmark for evaluating stable chart-to-values reading under increasing information load

## What Is Included

- `hf_viewer/{train,validation,test}/`
  - Hugging Face viewer-friendly split layout for the paper benchmark
- `metadata/formal_all18_432.jsonl`
  - canonical manifest for the 432 released samples
- `metadata/splits/{train,val,test}.jsonl`
  - split-specific manifests
- `images/`
  - rendered chart images for the 432 benchmark samples
- `tables/`
  - table-first ground-truth CSV files
- `programs/`
  - chart program / generation metadata JSON files
- `objects/`
  - object-level chart metadata JSON files
- `evaluation/`
  - truth files used by the formal paper workflow
- `docs/chartbench_e_rebuilt_README.md`
  - codebase README snapshot
- `docs/chartbench_e_paper.pdf`
  - paper PDF snapshot bundled with this release

## Release Summary

- Samples: `432`
- Chart families: `18`
- Samples per chart family: `24`
- Difficulty tiers:
  - `L1`: `144`
  - `L2`: `144`
  - `L3`: `144`
- Splits:
  - `train`: `289`
  - `val`: `50`
  - `test`: `93`

## Data Fields

The public viewer layout in `hf_viewer/{train,validation,test}/metadata.jsonl` keeps a slimmer schema:

- `sample_id`
- `chart_id`
- `chart_type`
- `chart_group`
- `image_path`
- `values`
- `n_series`
- `n_points`
- `difficulty_tier`
- `difficulty_components`
- `challenge_tags`
- `source_kind`
- `layout_family`
- `split`
- `suite_id`
- `annotation`
- `axis_scale`
- `x_labels`
- `series_labels`
- `candidate_pool`
- `table_path`
- `program_path`
- `object_path`

The canonical audit manifest remains available in `metadata/formal_all18_432.jsonl`.

## Notes

- This is the benchmark release cited by the manuscript.
- Code, prompt templates, evaluation scripts, and generation utilities live in:
  - `https://github.com/luojueling/ChartBench-E`
- The broader 7,506-sample released pool is published separately to keep this benchmark repo stable and viewer-friendly.

## Suggested Citation

```bibtex
@misc{chartbenche2026benchmark,
  title        = {ChartBench-E Benchmark 432},
  author       = {Hui Long and Jueling Luo and Si Xie and Haoli Yu and Juncheng Ye and Xutao Deng},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/xiaoluo11/ChartBench-E}},
  note         = {Frozen 432-sample benchmark release used in the manuscript}
}
```

## License

This release contains synthetic benchmark artifacts assembled for research evaluation. A project-level license statement should be added or updated before wider redistribution beyond the current release workflow.