---
configs:
- config_name: default
  data_files:
  - split: train
    path: data/repositories.parquet
- config_name: relations
  data_files:
  - split: train
    path: data/relations.parquet
---

## Snapshot status

Repositories: 5645805. Relations: 2296847.
Coverage is complete; see `manifest.json` for scan details.

# hf-ml

`hf-ml` is a machine learning repository registry built from Hugging Face Hub
metadata. The `default` configuration contains one row per observed model,
dataset, or Space repository. The `relations` configuration contains explicit
links between repositories, including derived or fine-tuned artifacts when the
source metadata supports that relationship.

Each snapshot includes a coverage manifest recording repository types, scan
windows, pagination, and any incomplete or failed pages. Coverage is therefore
measurable and auditable; this initial dataset does not claim exhaustive Hub
coverage. The registry stores metadata and relationships only. It does not copy
model weights, dataset payloads, or Space artifacts.

## Relationship semantics

For lineage relationships named `finetune`, `adapter`, `merge`, `quantized`, or
`base_model`, `source_key` identifies the child or derived model and
`target_key` identifies its parent or base model. A generic `base_model`
relationship is used when the metadata declares a base model but does not make
the more specific relationship type explicit. `target_exists: false` means the
referenced repository was not present in that snapshot; it does not invalidate
the relationship declaration.

Repository metadata and README content remain attributable to their original
Hub authors. See the manifest for the snapshot timestamp, source queries, and
file checksums.