# Hugging Face Eval Results

Real5-OmniDocBench defines six Hub leaderboard tasks in the repository root
`eval.yaml`:

- `overall`
- `scanning`
- `warping`
- `screen_photography`
- `illumination`
- `skew`

These correspond to the primary scores shown in the benchmark leaderboard. All
six are higher-is-better metrics.

## Model-Side Submission Format

Model repositories should add a file like:

```text
.eval_results/real5_omnidocbench.yaml
```

Use `examples/model_eval_results.yaml` as the template.

Each YAML list item reports one task:

```yaml
- dataset:
    id: PaddlePaddle/Real5-OmniDocBench
    task_id: overall
  value: 93.19
  date: "2026-05-28"
  source:
    url: https://huggingface.co/datasets/PaddlePaddle/Real5-OmniDocBench
    name: Real5-OmniDocBench Leaderboard
    user: PaddlePaddle
```

The public benchmark repository is used as the result source. Do not link
private or internal leaderboard data files. Add `notes` only when a result
needs an explanatory footnote, because the Hub renders notes with an asterisk.

## Generate Files From Current README Data

The Space leaderboard data must already be generated from the external source
file:

```bash
python leaderboard/scripts/build_from_source_data.py
```

Generate one YAML file per method:

```bash
python hf_eval_results/scripts/generate_model_eval_results.py
```

Generate only one method:

```bash
python hf_eval_results/scripts/generate_model_eval_results.py --method "PaddleOCR-VL-1.6"
```

Validate generated or submitted YAML:

```bash
python hf_eval_results/scripts/validate_eval_results.py hf_eval_results/generated
```

## Important Limitation

The Hub `.eval_results` schema stores one numeric `value` per dataset task.
This repository uses the six primary task scores for Hub-native submissions.
The richer scenario-level metrics remain available in the Static Space
leaderboard.