README.md

---
license: mit
task_categories:
- text-generation
language:
- en
tags:
- pcb
- schematic-design
- eda
- code-synthesis
- llm-benchmark
- hardware-design
- skidl
pretty_name: PCBSchemaGen Benchmarks (PCBBench + Open-Schematics-Eval)
size_categories:
- n<1K
configs:
- config_name: pcbbench
  data_files:
  - split: test
    path: pcbbench/pcbbench.jsonl
- config_name: open_schematics
  data_files:
  - split: test
    path: open_schematics/open_schematics.jsonl
---

# PCBSchemaGen Benchmarks

Two benchmark suites for **LLM-driven PCB schematic synthesis**, from the paper
[*PCBSchemaGen: Reward-Guided LLM Code Synthesis for Printed Circuit Board (PCB) Schematic Design with Structured Verification*](https://arxiv.org/abs/2602.00510).

Correctness in this domain is **not** defined by unit tests: there are no per-task golden references,
and SPICE does not validate schematic-level correctness. Instead, each task is scored by a
**deterministic structural verifier** against real-IC pin- and topology-level constraints
(no LLM, no API key, no cached results). The verifier and knowledge graphs live in the
[GitHub repository](https://github.com/HZou9/PCBSchemaGen_v2).

## Suites

| Config | Tasks | Easy / Medium / Hard | Real ICs | Domains |
|---|---|---|---|---|
| `pcbbench` | 62 | 17 / 28 / 17 | 41 commercial | 22 |
| `open_schematics` | 165 | 67 / 63 / 35 | 439 commercial | 22 |
| **Total** | **227** | — | 480 | 22 |

## Load

```python
from datasets import load_dataset

pcbbench = load_dataset("Hzou9/PCBSchemaGen-Benchmarks", "pcbbench", split="test")
ose      = load_dataset("Hzou9/PCBSchemaGen-Benchmarks", "open_schematics", split="test")
print(pcbbench[0])
```

## Fields

Both suites share these 10 core fields:

| Field | Description |
|---|---|
| `id` | Task identifier |
| `level` | Difficulty: `Easy` / `Medium` / `Hard` |
| `type` | Circuit type (e.g. Sensing, Power, Digital) |
| `task` | Natural-language design specification |
| `input_nodes` | Required input net(s) |
| `output_nodes` | Required output net(s) |
| `input_voltage` | Input voltage / range |
| `output_voltage` | Output voltage / range |
| `components` | Required component set (real ICs and/or passives) |
| `sub_module_name` | Canonical sub-module name for the target circuit |

`open_schematics` additionally carries two optional ground-truth fields, present on a subset
(20/165) of tasks and `null` elsewhere:

| Field | Description |
|---|---|
| `n_gt_components` | Ground-truth component count for the reference schematic |
| `gt_component_types` | Number of distinct ground-truth component types |

## Evaluation

A candidate is a SKiDL (Python) schematic-generation program. Run it, then score the resulting
schematic with the deterministic 5-layer verifier in the GitHub repo (`framework/topo/`), which
checks pin-level assignments and topology against constraints induced from real IC datasheets.
See the repo README for the runner and per-task configuration.

## License

MIT — see `LICENSE`. © 2026 Huanghaohe Zou, Peng Han, Emad Nazerian, Mafu Zhang, Zhicheng Guo, Alex Q. Huang.

## Citation

If you use PCBBench or Open-Schematics-Eval, please cite:

```bibtex
@article{zou2026pcbschemagen,
  title   = {PCBSchemaGen: Reward-Guided LLM Code Synthesis for Printed Circuit Board (PCB) Schematic Design with Structured Verification},
  author  = {Zou, Huanghaohe and Han, Peng and Nazerian, Emad and Zhang, Mafu and Guo, Zhicheng and Huang, Alex Q.},
  journal = {arXiv preprint arXiv:2602.00510},
  year    = {2026},
  url     = {https://arxiv.org/abs/2602.00510}
}
```