ds-cache-cleaner

Clean up cached data from ML/data science libraries.

Supported Caches

• HuggingFace Models - ~/.cache/huggingface/hub (models)
• HuggingFace Datasets (Hub) - ~/.cache/huggingface/hub (datasets)
• Transformers - ~/.cache/huggingface/transformers
• HF Datasets - ~/.cache/huggingface/datasets
• ir_datasets - ~/.ir_datasets
• datamaestro (cache) - ~/datamaestro/cache (partial downloads, processing)
• datamaestro (data) - ~/datamaestro/data (downloaded datasets)

Installation

pip install ds-cache-cleaner

Or with uv:

uv pip install ds-cache-cleaner

Usage

List caches

ds-cache-cleaner list

Show cache entries

ds-cache-cleaner show
ds-cache-cleaner show -c "HuggingFace Hub"

Clean caches

# Interactive mode
ds-cache-cleaner clean

# Clean specific cache
ds-cache-cleaner clean -c "HuggingFace Hub"

# Clean all without prompting
ds-cache-cleaner clean --all

# Dry run
ds-cache-cleaner clean --dry-run

Interactive TUI

ds-cache-cleaner tui

Library Integration

ML libraries can integrate with ds-cache-cleaner to provide rich metadata about their cached data. This enables better descriptions, accurate last-access times, and custom library-specific metadata.

Metadata Format

The metadata is stored in a ds-cache-cleaner/ folder inside each cache directory:

~/.cache/mylib/
├── ds-cache-cleaner/
│   ├── lock                    # Lock file for concurrent access
│   ├── information.json        # Cache info and parts list
│   └── part_models.json        # Entries for "models" part
└── ... (actual cache data)

Using the CacheRegistry API

from ds_cache_cleaner import CacheRegistry

# Initialize once for your library
registry = CacheRegistry(
    cache_path="~/.cache/mylib",
    library="mylib",
    description="My ML Library cache",
)

# Register a part (e.g., models, datasets)
registry.register_part("models", "Downloaded model weights")

# When downloading a new model
registry.register_entry(
    part="models",
    path="bert-base",  # relative path within cache
    description="BERT base model",
    size=438_000_000,
)

# When accessing an existing entry (updates last_access time)
registry.touch("models", "bert-base")

# When deleting an entry (removes from metadata)
registry.remove("models", "bert-base")

Custom Metadata

You can store library-specific metadata using the metadata parameter:

# Store custom metadata when registering an entry
registry.register_entry(
    part="models",
    path="bert-base-uncased",
    description="BERT base uncased model",
    size=438_000_000,
    metadata={
        "model_type": "bert",
        "revision": "main",
        "tags": ["encoder", "uncased", "english"],
        "framework": "pytorch",
    },
)

# Retrieve entry with its metadata
entry = registry.get_entry("models", "bert-base-uncased")
if entry:
    print(entry.metadata)  # {"model_type": "bert", ...}

Full API Reference

CacheRegistry

Parts Accessor

The parts property provides a convenient attribute-style API:

# Instead of:
registry.register_entry("models", entry)
registry.get_entry("models", "bert")
registry.list_entries("models")

# You can use:
registry.parts.models.register(entry)
registry.parts.models.get("bert")
registry.parts.models.list()

PartAccessor methods:

Data Classes

The metadata system uses standard Python dataclasses with pydantic validation:

from ds_cache_cleaner import EntryMetadata, PartInfo

# EntryMetadata fields
@dataclass
class EntryMetadata:
    path: str                           # Relative path within cache (required)
    description: str = ""               # Human-readable description
    created: datetime | None = None     # When the entry was created
    last_access: datetime | None = None # Last access time
    size: int | None = None             # Size in bytes
    metadata: dict[str, Any] = {}       # Library-specific metadata

# PartInfo fields
@dataclass
class PartInfo:
    name: str          # Part name (e.g., "models")
    description: str = ""  # Human-readable description

Custom Entry Classes

For type-safe entries, subclass EntryMetadata with your own fields. Extra fields are automatically serialized into the metadata dict and reconstructed on read:

from dataclasses import dataclass, field
from ds_cache_cleaner import CacheRegistry, EntryMetadata

@dataclass
class ModelEntry(EntryMetadata):
    """Custom entry for ML models."""
    model_type: str = ""
    revision: str = ""
    tags: list[str] = field(default_factory=list)

# Create registry with custom entry type
registry = CacheRegistry(
    cache_path="~/.cache/mylib",
    library="mylib",
    entry_types={"models": ModelEntry},  # Map part name to entry type
)

registry.register_part("models", "Model weights")

# Register using custom entry instance
entry = ModelEntry(
    path="bert-base",
    description="BERT model",
    model_type="bert",
    revision="v1.0",
    tags=["encoder", "english"],
)
registry.register_entry("models", entry)

# Get entry returns the correct type
model = registry.get_entry("models", "bert-base")
assert isinstance(model, ModelEntry)
print(model.model_type)  # "bert"
print(model.tags)        # ["encoder", "english"]

# List entries also returns custom types
for entry in registry.list_entries("models"):
    print(f"{entry.path}: {entry.model_type}")

The JSON format remains backward compatible - extra fields are stored in the metadata dict:

{
  "path": "bert-base",
  "description": "BERT model",
  "metadata": {
    "model_type": "bert",
    "revision": "v1.0",
    "tags": ["encoder", "english"]
  }
}

Development

# Install with dev dependencies
pip install -e ".[dev]"

# Run tests
hatch run test

# Lint
hatch run lint:check

# Format
hatch run lint:fix

License

MIT