# Hugging Face quickstart

The expanded `repo/` folder becomes the root of `PureOne/AELYRION-NOULITH-III` after upload.

## Download a deposited snapshot

```bash
python -m pip install "huggingface_hub>=0.35,<3"
```

```python
from huggingface_hub import snapshot_download
snapshot_download(
    repo_id="PureOne/AELYRION-NOULITH-III",
    repo_type="dataset",
    local_dir="AELYRION-NOULITH-III",
)
```

This remote example requires the prepared repository to have been uploaded. For a pinned
reproduction, pass `revision` set to the actual deposited commit hash.

## Run the exact example

From the downloaded repository root:

```bash
python -m pip install -e .
python hub_tools/quickstart_exact.py
```

Defaults: six channels, 41 scalar components, twin rank 60 of 64, four synchronizing twin
constraints, and one edge-scale class. A generic visible witness is verified in the original
sparse algebra. All certificate coefficients are exact rationals.

`--m` selects another channel count. Above eight, the quickstart omits the exponential
independent rank calculation and reports that omission.

## Run tests and exhaustive finite checks

```bash
python -m pip install -r requirements.txt
python -m pytest -q
python scripts/exhaustive_v3.py
```

The supplied suite has 1,028 cases. The graph script performs two distinct exhaustive
six-channel enumerations and finds 810 minimum-size synchronizing families. These remain
finite computational checks, separate from the general written proofs.

## Recreate the notebooks and PDF

```bash
python -m pip install -r requirements-notebook.txt
python scripts/reproduce.py --notebook --skip-pdf
```

Both original notebooks are already executed. The command regenerates data, validation
records, and notebook outputs, intentionally changing the delivered bytes.
For the PDF, install TeX Live or MiKTeX with `pdflatex` and `bibtex`, then run:

```bash
python scripts/build_manuscript.py
```

The delivered PDF has 67 pages and can be read without rebuilding.

## Load an individual table

```bash
python -m pip install "datasets>=4,<6"
```

```python
from datasets import load_dataset
table = load_dataset(
    "PureOne/AELYRION-NOULITH-III",
    "v3_exact_rank_strata",
    split="train",
)
```

See [DATA_DICTIONARY.md](DATA_DICTIONARY.md) for row semantics. All configurations have a
single logical split; there is no held-out ML benchmark.

## Integrity

```bash
python hub_tools/verify_release.py
```

After intentional edits:

```bash
python hub_tools/prepare_viewer_data.py
python hub_tools/verify_release.py --write
```

The manifest verifies prepared-file integrity, not authorship or mathematical truth.
The original ZIP has its own unchanged internal manifest.

## API

Channel labels are zero-based; coefficient keys are ascending pairs. Use exact SymPy
rationals. `certify_scalar` enforces m>=6 and abs(q_ij+1)<1/20 and rejects floating-point
coefficients. [API_REFERENCE.md](API_REFERENCE.md) lists the existing entry points.

## Check the Dataset configuration locally

After dependency installation, this uses local files and offline Hub settings:

```bash
python -m pip install -r requirements-hub-validation.txt
python hub_tools/validate_dataset_configs.py
```