---
pretty_name: PhysAlign
language:
- en
- zh
license: other
task_categories:
- question-answering
- visual-question-answering
tags:
- physics
- multimodal
- benchmark
- evaluation
- image
size_categories:
- 1K<n<10K
configs:
- config_name: raw
  data_files:
  - split: development
    path: data/raw/development.jsonl
  - split: test
    path: data/raw/test.jsonl
- config_name: gold
  data_files:
  - split: development
    path: data/gold/development.jsonl
  - split: test
    path: data/gold/test.jsonl
---

# PhysAlign

PhysAlign is a bilingual multimodal physics benchmark for testing whether a model can make local observations and bind textual or visual references to the correct candidate entity. It does **not** ask the evaluated model to solve the original examination problem.

PhysAlign was constructed from selected examples in previously released public benchmarks, followed by source adaptation, probe generation, model-based review, cross-source exact-input deduplication, and split checks.



## Dataset summary

| Source component | Mother problems | Raw | Gold |
|---|---:|---:|---:|
| local (SeePhys + LiveK12Bench) | 458 | 1,679 | 211 |
| OlympiadBench | 209 | 669 | 97 |
| PhysElite | 220 | 792 | 104 |
| PhyX | 16 | 40 | 35 |
| GAOKAO-MM | 23 | 38 | 26 |
| expansion_v3 (SeePhys) | 60 | 123 | 80 |
| **Total** | **986** | **3,341** | **553** |

| Split | Mother problems | Raw | Gold |
|---|---:|---:|---:|
| development | 79 | 208 | 90 |
| test | 907 | 3,133 | 463 |

Raw interface counts:

- `T02-single`: 2,407
- `T03-image`: 514
- `T03-text`: 420

Languages:

- Raw: 3,129 English and 212 Chinese instances.
- Gold: 458 English and 95 Chinese instances.

## Released files

```text
.
|-- README.md
|-- LICENSE.md
|-- THIRD_PARTY_NOTICES.md
|-- release.json
|-- MANIFEST.sha256
|-- data/
|   |-- raw/
|   |   |-- development.jsonl
|   |   `-- test.jsonl
|   `-- gold/
|       |-- development.jsonl
|       `-- test.jsonl
`-- public/
    |-- qa_raw.json
    |-- qa_gold.json
    `-- images/
        |-- 00/
        |-- 01/
        |-- ...
        `-- ff/
```

- `public/qa_raw.json` and `public/qa_gold.json` are the canonical array-form benchmark inputs.
- `data/` contains equivalent JSONL copies split by condition and benchmark split for the Hugging Face Dataset Viewer.
- Images are sharded by the first two hexadecimal characters of their SHA-256 filename so that no Hub directory exceeds the file-count limit.
- Every attachment path is resolved relative to `public/`.
- `release.json` contains sanitized release statistics and provenance metadata. Service endpoint information has been removed.
- `MANIFEST.sha256` contains SHA-256 checksums for every released file except the manifest itself.

Scoring answers, source-key mappings, private membership records, internal review traces, and archived component releases are intentionally **not included**.

## Record schema

Each record contains:

| Field | Meaning |
|---|---|
| `condition` | `raw` or `gold` |
| `instance_id` | Stable public instance identifier |
| `logical_probe_id` | Identifier shared by related Raw/Gold conditions |
| `task_id` | `T02` or `T03` |
| `interface` | Expected response interface |
| `language` | `en` or `zh` |
| `split` | `development` or `test` |
| `schema_version` | Public schema version |
| `input.messages` | System and user messages for the evaluated model |
| `input.attachments` | Image metadata and paths relative to `public/` |

Interfaces:

- `T02-single`: return `{"referent":"E..."}`.
- `T03-image`: return `{"read":"...","owner":"E..."}`.
- `T03-text`: return `{"owner":"E..."}`.

The `read` value is evaluated with the matching PhysAlign normalizer. The evaluator does not perform additional physics inference or unit conversion for this field.

Gold is a supplementary input condition with local transcription information. It is not an additional independent question and it does not contain the hidden scoring answer.

## Loading

### Hugging Face Datasets

```python
from datasets import load_dataset

raw = load_dataset("Jetson888/PhysAlign", "raw")
gold = load_dataset("Jetson888/PhysAlign", "gold")

raw_dev = raw["development"]
raw_test = raw["test"]
```

The Dataset Viewer exposes attachment metadata as nested values. To access image files, download the repository snapshot as shown below.

### Download the complete benchmark

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

root = Path(snapshot_download(
    repo_id="Jetson888/PhysAlign",
    repo_type="dataset",
))

public_root = root / "public"
records = json.loads((public_root / "qa_raw.json").read_text(encoding="utf-8"))

first_attachment = records[0]["input"]["attachments"][0]
image_path = public_root / first_attachment["path"]
print(image_path)
```

CLI equivalent:

```bash
hf download Jetson888/PhysAlign --repo-type dataset --local-dir PhysAlign
```

## Evaluation guidance

1. Use the `development` split for integration and prompt-format debugging.
2. Send only the released public inputs to the evaluated model.
3. Aggregate or resample by mother problem rather than treating every probe as independent.
4. Report Raw and Gold separately.
5. Report `T02-single`, `T03-image`, and `T03-text` separately.
6. Use a matching version of the PhysAlign benchmark kit and its normalizers for scoring.

## Data status and limitations

- `human_reviewed=false` and `independent_human_quality_certified=false`.
- Exact duplicate public inputs across sources were collapsed.
- Near-duplicate reconciliation and historical development-sample reconciliation are incomplete.
- Some source-level or conversion-level findings remain unresolved; affected mother problems were quarantined where identified.
- Shared-image mothers were grouped for split-risk control but are not asserted to be semantic duplicates.
- The release should not be presented as a zero-error dataset or as an independently certified unseen-test benchmark.
- Original source data may carry additional limitations and usage conditions; see `THIRD_PARTY_NOTICES.md` and the upstream dataset cards.

## Sources and license

This is a composite dataset. The repository therefore uses `license: other` rather than applying one upstream license to all records. PhysAlign-authored packaging and metadata, and all third-party attributions and conditions, are described in [LICENSE.md](LICENSE.md) and [THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md).

The public inputs were derived from:

- [SeePhys](https://huggingface.co/datasets/SeePhys/SeePhys)
- [LiveK12Bench](https://huggingface.co/datasets/Shawn-wxh/livek12bench)
- [OlympiadBench](https://github.com/OpenBMB/OlympiadBench)
- [PhysElite](https://huggingface.co/datasets/physelite/PhysElite)
- [PhyX](https://github.com/killthefullmoon/PhyX)
- [GAOKAO-MM](https://github.com/OpenMOSS/GAOKAO-MM)

## Citation

If you use PhysAlign, cite this repository and the applicable upstream datasets listed in `THIRD_PARTY_NOTICES.md`.

```bibtex
@dataset{physalign2026,
  title     = {PhysAlign: A Multimodal Physics Local-Observation and Role-Binding Benchmark},
  author    = {PhysAlign Contributors},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/Jetson888/PhysAlign}
}
```

---

## 中文说明

PhysAlign 用于评测多模态模型能否正确完成局部观察、可见文本转录和指代/角色绑定，而不是要求模型解答原始物理题。公开仓库只包含模型输入与图片；评分答案、映射、审核记录及历史封存副本均不会公开。