glitchlings 0.4.0__cp310-cp310-macosx_11_0_universal2.whl → 0.4.2__cp310-cp310-macosx_11_0_universal2.whl

glitchlings/dlc/huggingface.py

@@ -3,21 +3,15 @@
 from __future__ import annotations
 
 from collections.abc import Iterable, Sequence
-from typing import Any
+from typing import Any, cast
 
-try:  # pragma: no cover - optional dependency is required at runtime
-    from datasets import Dataset as _DatasetsDataset
-except ModuleNotFoundError as _datasets_error:  # pragma: no cover - optional dependency
-    _DatasetsDataset = None  # type: ignore[assignment]
-else:
-    _datasets_error = None
-
-from ..zoo import Gaggle, Glitchling, summon
+from ..compat import datasets, get_datasets_dataset, require_datasets
+from ..util.adapters import coerce_gaggle
+from ..zoo import Gaggle, Glitchling
 
 
 def _normalise_columns(column: str | Sequence[str]) -> list[str]:
     """Normalise a column specification to a list."""
-
     if isinstance(column, str):
         return [column]
 
@@ -27,20 +21,6 @@ def _normalise_columns(column: str | Sequence[str]) -> list[str]:
     return normalised
 
 
-def _as_gaggle(glitchlings: Glitchling | Gaggle | str | Iterable[str | Glitchling], seed: int) -> Gaggle:
-    """Coerce any supported glitchling specification into a :class:`Gaggle`."""
-
-    if isinstance(glitchlings, Gaggle):
-        return glitchlings
-
-    if isinstance(glitchlings, (Glitchling, str)):
-        resolved: Iterable[str | Glitchling] = [glitchlings]
-    else:
-        resolved = glitchlings
-
-    return summon(list(resolved), seed=seed)
-
-
 def _glitch_dataset(
     dataset: Any,
     glitchlings: Glitchling | Gaggle | str | Iterable[str | Glitchling],
@@ -48,23 +28,28 @@ def _glitch_dataset(
     *,
     seed: int = 151,
 ) -> Any:
-    """Internal helper implementing :meth:`Dataset.glitch`."""
-
+    """Apply glitchlings to the provided dataset columns."""
     columns = _normalise_columns(column)
-    gaggle = _as_gaggle(glitchlings, seed=seed)
+    gaggle = coerce_gaggle(glitchlings, seed=seed)
     return gaggle.corrupt_dataset(dataset, columns)
 
 
 def _ensure_dataset_class() -> Any:
     """Return the Hugging Face :class:`~datasets.Dataset` patched with ``.glitch``."""
-
-    if _DatasetsDataset is None:  # pragma: no cover - datasets is an install-time dependency
-        message = "datasets is not installed"
-        raise ModuleNotFoundError(message) from _datasets_error
-
-    if getattr(_DatasetsDataset, "glitch", None) is None:
-
-        def glitch(  # type: ignore[override]
+    dataset_cls = get_datasets_dataset()
+    if dataset_cls is None:  # pragma: no cover - datasets is an install-time dependency
+        require_datasets("datasets is not installed")
+        dataset_cls = get_datasets_dataset()
+        if dataset_cls is None:
+            message = "datasets is not installed"
+            error = datasets.error
+            if error is not None:
+                raise ModuleNotFoundError(message) from error
+            raise ModuleNotFoundError(message)
+
+    if getattr(dataset_cls, "glitch", None) is None:
+
+        def glitch(
             self: Any,
             glitchlings: Glitchling | Gaggle | str | Iterable[str | Glitchling],
             *,
@@ -73,24 +58,24 @@ def _ensure_dataset_class() -> Any:
             **_: Any,
         ) -> Any:
             """Return a lazily corrupted copy of the dataset."""
-
             return _glitch_dataset(self, glitchlings, column, seed=seed)
 
-        setattr(_DatasetsDataset, "glitch", glitch)
+        setattr(dataset_cls, "glitch", glitch)
 
-    return _DatasetsDataset
+    return cast(type[Any], dataset_cls)
 
 
 def install() -> None:
     """Monkeypatch the Hugging Face :class:`~datasets.Dataset` with ``.glitch``."""
-
     _ensure_dataset_class()
 
 
-if _DatasetsDataset is not None:
+Dataset: type[Any] | None
+_DatasetAlias = get_datasets_dataset()
+if _DatasetAlias is not None:
     Dataset = _ensure_dataset_class()
 else:  # pragma: no cover - datasets is an install-time dependency
-    Dataset = None  # type: ignore[assignment]
+    Dataset = None
 
 
 __all__ = ["Dataset", "install"]

glitchlings/dlc/prime.py

@@ -4,79 +4,60 @@ from __future__ import annotations
 
 from collections.abc import Iterable, Sequence
 from enum import Enum
-from typing import Any, Callable
+from typing import Any, Callable, Protocol, cast
 
-import verifiers as vf
+from ..compat import require_datasets, require_jellyfish, require_verifiers
+from ..util.adapters import coerce_gaggle
+from ..zoo import Gaggle, Glitchling, Mim1c, Typogre
+from ._shared import resolve_columns as _resolve_columns_shared
+from ._shared import resolve_environment as _resolve_environment_shared
 
-from jellyfish import damerau_levenshtein_distance
 
-try:
-    from .huggingface import Dataset
-except ModuleNotFoundError:  # pragma: no cover - optional dependency
-    Dataset = object  # type: ignore[assignment]
-else:
-    if Dataset is None:  # pragma: no cover - optional dependency
-        Dataset = object  # type: ignore[assignment]
+class VerifierEnvironment(Protocol):
+    """Minimal interface for verifiers environments."""
 
-from ..zoo import Gaggle, Glitchling, Mim1c, Typogre, summon
+    dataset: Any
 
 
-def _resolve_environment(env: str | vf.Environment) -> vf.Environment:
-    """Return a fully-instantiated verifier environment."""
-
-    if isinstance(env, str):
-        env = vf.load_environment(env)
+class VerifierSingleTurnEnv(Protocol):
+    """Minimal interface for single-turn verifier environments."""
 
-    if not isinstance(env, vf.Environment):
-        raise TypeError("Invalid environment type")
+    dataset: Any
+    rubric: Any
 
-    return env
 
+vf = require_verifiers("verifiers is not installed; install glitchlings[prime]")
+_jellyfish = require_jellyfish("jellyfish is not installed; install glitchlings[prime]")
+damerau_levenshtein_distance = _jellyfish.damerau_levenshtein_distance
 
-def _resolve_columns(dataset: Dataset, columns: Sequence[str] | None) -> list[str]:
-    """Identify which dataset columns should be corrupted."""
-
-    available = set(dataset.column_names)
+try:
+    from .huggingface import Dataset as _HuggingFaceDataset
+except ModuleNotFoundError:  # pragma: no cover - optional dependency
+    _HuggingFaceDataset = None
+else:
+    if _HuggingFaceDataset is None:  # pragma: no cover - optional dependency
+        _HuggingFaceDataset = None
 
-    if columns is not None:
-        missing = sorted(set(columns) - available)
-        if missing:
-            missing_str = ", ".join(missing)
-            raise ValueError(f"Columns not found in dataset: {missing_str}")
-        return list(columns)
+Dataset: type[Any]
+if _HuggingFaceDataset is None:
+    Dataset = object
+else:
+    Dataset = _HuggingFaceDataset
 
-    for candidate in ("prompt", "question"):
-        if candidate in available:
-            return [candidate]
 
-    try:
-        dataset_length = len(dataset)  # type: ignore[arg-type]
-    except TypeError:
-        preview_rows: list[dict[str, Any]]
-        take_fn = getattr(dataset, "take", None)
-        if callable(take_fn):
-            preview_rows = list(take_fn(1))
-        else:
-            iterator = iter(dataset)
-            try:
-                first_row = next(iterator)
-            except StopIteration:
-                preview_rows = []
-            else:
-                preview_rows = [first_row]
-        sample = dict(preview_rows[0]) if preview_rows else {}
-    else:
-        sample = dataset[0] if dataset_length else {}
-    inferred = [
-        name
-        for name in dataset.column_names
-        if isinstance(sample.get(name), str)
-    ]
+def _resolve_environment(env: str | VerifierEnvironment) -> VerifierEnvironment:
+    """Return a fully-instantiated verifier environment."""
+    resolved = _resolve_environment_shared(
+        env,
+        loader=vf.load_environment,
+        environment_type=cast(type[Any], vf.Environment),
+    )
+    return cast(VerifierEnvironment, resolved)
 
-    if inferred:
-        return inferred
 
-    raise ValueError("Unable to determine which dataset columns to corrupt.")
+def _resolve_columns(dataset: Any, columns: Sequence[str] | None) -> list[str]:
+    """Identify which dataset columns should be corrupted."""
+    return _resolve_columns_shared(dataset, columns)
 
 
 class Difficulty(Enum):
@@ -90,12 +71,11 @@ class Difficulty(Enum):
 
 
 def tutorial_level(
-    env: vf.Environment | str,
+    env: VerifierEnvironment | str,
     seed: int = 151,
     difficulty: Difficulty = Difficulty.Normal,
-) -> vf.Environment:
+) -> VerifierEnvironment:
     """Create a low-corruption environment using tuned defaults."""
-
     tuned_mim1c = Mim1c(rate=0.01 * difficulty.value)
     tuned_typogre = Typogre(rate=0.025 * difficulty.value)
 
@@ -107,28 +87,19 @@ def tutorial_level(
 
 
 def load_environment(
-    env: str | vf.Environment,
+    env: str | VerifierEnvironment,
     glitchlings: Iterable[str | Glitchling] | Glitchling | str | Gaggle | None = None,
     *,
     seed: int = 151,
     columns: Sequence[str] | None = None,
-) -> vf.Environment:
+) -> VerifierEnvironment:
     """Load an environment and optionally corrupt it with glitchlings."""
-
     environment = _resolve_environment(env)
 
     if glitchlings is None:
         return environment
 
-    if isinstance(glitchlings, Gaggle):
-        gaggle = glitchlings
-    else:
-        if isinstance(glitchlings, (Glitchling, str)):
-            resolved = [glitchlings]
-        else:
-            resolved = list(glitchlings)
-
-        gaggle = summon(resolved, seed=seed)
+    gaggle = coerce_gaggle(glitchlings, seed=seed)
 
     dataset = environment.dataset
     corrupt_columns = _resolve_columns(dataset, columns)
@@ -142,21 +113,11 @@ def _as_gaggle(
     seed: int,
 ) -> Gaggle:
     """Coerce any supported glitchling specification into a :class:`Gaggle`."""
-
-    if isinstance(glitchlings, Gaggle):
-        return glitchlings
-
-    if isinstance(glitchlings, (Glitchling, str)):
-        resolved: Iterable[str | Glitchling] = [glitchlings]
-    else:
-        resolved = glitchlings
-
-    return summon(list(resolved), seed=seed)
+    return coerce_gaggle(glitchlings, seed=seed)
 
 
 def _extract_completion_text(completion: Any) -> str:
     """Normalise a completion payload into a plain string."""
-
     if isinstance(completion, str):
         return completion
 
@@ -175,11 +136,10 @@ def symmetric_damerau_levenshtein_similarity(
     answer: str,
 ) -> float:
     """Return ``1 - (distance / max_len)`` using Damerau-Levenshtein distance."""
-
     completion_text = _extract_completion_text(completion)
     target = answer or ""
     denominator = max(len(completion_text), len(target), 1)
-    distance = damerau_levenshtein_distance(completion_text, target)
+    distance = cast(int, damerau_levenshtein_distance(completion_text, target))
     score = 1.0 - (distance / denominator)
     return max(0.0, min(1.0, score))
 
@@ -199,32 +159,34 @@ def echo_chamber(
     reward_function: Callable[..., float] | None = None,
     split: str | None = None,
     **load_dataset_kwargs: Any,
-) -> vf.Environment:
+) -> VerifierSingleTurnEnv:
     """Create an Echo Chamber Prime environment from a Hugging Face dataset column.
 
     Args:
         dataset_id: Identifier of the Hugging Face dataset to load.
         column: Name of the column whose text should be glitched.
         glitchlings: Glitchling specifiers that will corrupt the prompts.
-        seed: RNG seed forwarded to :func:`summon`.
+        seed: RNG seed forwarded to :func:`glitchlings.util.adapters.coerce_gaggle`.
         instructions: System instructions supplied to the environment prompts.
         reward_function: Optional callable used to score completions. Defaults to
             :func:`symmetric_damerau_levenshtein_similarity` when omitted.
         split: Optional dataset split to load.
         **load_dataset_kwargs: Extra keyword arguments forwarded to
             :func:`datasets.load_dataset`.
-    """
 
-    try:
-        from datasets import Dataset as HFDataset, DatasetDict, load_dataset
-    except ModuleNotFoundError as exc:  # pragma: no cover - optional dependency
+    """
+    datasets_module = require_datasets("datasets is required to build an echo chamber")
+    load_dataset = getattr(datasets_module, "load_dataset", None)
+    if load_dataset is None:  # pragma: no cover - defensive
         message = "datasets is required to build an echo chamber"
-        raise ModuleNotFoundError(message) from exc
+        raise ModuleNotFoundError(message)
 
-    hf_dataset: HFDataset | DatasetDict
+    dataset_dict_cls = getattr(datasets_module, "DatasetDict", dict)
+
+    hf_dataset: Any
     if split is None:
         hf_dataset = load_dataset(dataset_id, **load_dataset_kwargs)
-        if isinstance(hf_dataset, DatasetDict):
+        if isinstance(hf_dataset, dataset_dict_cls):
             try:
                 hf_dataset = next(iter(hf_dataset.values()))
             except StopIteration as exc:  # pragma: no cover - defensive
@@ -232,10 +194,8 @@ def echo_chamber(
     else:
         hf_dataset = load_dataset(dataset_id, split=split, **load_dataset_kwargs)
 
-    if isinstance(hf_dataset, DatasetDict):
-        raise ValueError(
-            "Specify which split to use when the dataset loads as a DatasetDict."
-        )
+    if isinstance(hf_dataset, dataset_dict_cls):
+        raise ValueError("Specify which split to use when the dataset loads as a DatasetDict.")
 
     filtered_dataset = hf_dataset.filter(
         lambda row: row.get(column) is not None,
@@ -259,7 +219,7 @@ def echo_chamber(
     )
 
     try:
-        dataset_length = len(base_dataset)  # type: ignore[arg-type]
+        dataset_length = len(base_dataset)
     except TypeError:
         preview_rows: list[dict[str, Any]]
         take_fn = getattr(base_dataset, "take", None)
@@ -288,4 +248,7 @@ def echo_chamber(
 
     rubric_func = reward_function or symmetric_damerau_levenshtein_similarity
     rubric = vf.Rubric(funcs=[rubric_func], weights=[1.0])
-    return vf.SingleTurnEnv(dataset=glitched_dataset, rubric=rubric)
+    return cast(
+        VerifierSingleTurnEnv,
+        vf.SingleTurnEnv(dataset=glitched_dataset, rubric=rubric),
+    )

glitchlings/lexicon/__init__.py

@@ -2,13 +2,16 @@
 
 from __future__ import annotations
 
+import random
 from abc import ABC, abstractmethod
 from hashlib import blake2s
-import random
+from pathlib import Path
 from typing import Callable, Iterable
 
 from glitchlings.config import get_config
 
+from ._cache import CacheEntries, CacheSnapshot
+
 
 class Lexicon(ABC):
     """Abstract interface describing synonym lookup backends.
@@ -19,6 +22,7 @@ class Lexicon(ABC):
         Optional integer used to derive deterministic random number generators
         for synonym sampling. Identical seeds guarantee reproducible results for
         the same word/part-of-speech queries.
+
     """
 
     def __init__(self, *, seed: int | None = None) -> None:
@@ -27,17 +31,14 @@ class Lexicon(ABC):
     @property
     def seed(self) -> int | None:
         """Return the current base seed used for deterministic sampling."""
-
         return self._seed
 
     def reseed(self, seed: int | None) -> None:
         """Update the base seed driving deterministic synonym sampling."""
-
         self._seed = seed
 
     def _derive_rng(self, word: str, pos: str | None) -> random.Random:
         """Return an RNG derived from the base seed, word, and POS tag."""
-
         seed_material = blake2s(digest_size=8)
         seed_material.update(word.lower().encode("utf8"))
         if pos is not None:
@@ -51,7 +52,6 @@ class Lexicon(ABC):
         self, values: Iterable[str], *, limit: int, word: str, pos: str | None
     ) -> list[str]:
         """Return up to ``limit`` values sampled deterministically."""
-
         if limit <= 0:
             return []
 
@@ -65,28 +65,40 @@ class Lexicon(ABC):
         return [items[index] for index in indices]
 
     @abstractmethod
-    def get_synonyms(
-        self, word: str, pos: str | None = None, n: int = 5
-    ) -> list[str]:
+    def get_synonyms(self, word: str, pos: str | None = None, n: int = 5) -> list[str]:
         """Return up to ``n`` synonyms for ``word`` constrained by ``pos``."""
 
     def supports_pos(self, pos: str | None) -> bool:
         """Return ``True`` when the backend can service ``pos`` queries."""
-
         return True
 
     def __repr__(self) -> str:  # pragma: no cover - trivial representation
         return f"{self.__class__.__name__}(seed={self._seed!r})"
 
 
-from .graph import GraphLexicon
-from .metrics import (
+class LexiconBackend(Lexicon):
+    """Extended lexicon interface that supports cache persistence."""
+
+    Cache = CacheEntries
+
+    @classmethod
+    @abstractmethod
+    def load_cache(cls, path: str | Path) -> CacheSnapshot:
+        """Return a validated cache snapshot loaded from ``path``."""
+
+    @abstractmethod
+    def save_cache(self, path: str | Path | None = None) -> Path | None:
+        """Persist the backend cache to ``path`` and return the destination."""
+
+
+from .graph import GraphLexicon  # noqa: E402
+from .metrics import (  # noqa: E402
     compare_lexicons,
     coverage_ratio,
     mean_cosine_similarity,
     synonym_diversity,
 )
-from .vector import VectorLexicon, build_vector_cache
+from .vector import VectorLexicon, build_vector_cache  # noqa: E402
 
 try:  # pragma: no cover - optional dependency
     from .wordnet import WordNetLexicon
@@ -97,24 +109,19 @@ except Exception:  # pragma: no cover - triggered when nltk unavailable
 _BACKEND_FACTORIES: dict[str, Callable[[int | None], Lexicon | None]] = {}
 
 
-def register_backend(
-    name: str, factory: Callable[[int | None], Lexicon | None]
-) -> None:
+def register_backend(name: str, factory: Callable[[int | None], Lexicon | None]) -> None:
     """Register ``factory`` for ``name`` so it can be selected via config."""
-
     normalized = name.lower()
     _BACKEND_FACTORIES[normalized] = factory
 
 
 def unregister_backend(name: str) -> None:
     """Remove a previously registered backend."""
-
     _BACKEND_FACTORIES.pop(name.lower(), None)
 
 
 def available_backends() -> list[str]:
     """Return the names of registered lexicon factories."""
-
     return sorted(_BACKEND_FACTORIES)
 
 
@@ -155,7 +162,6 @@ register_backend("wordnet", _wordnet_backend)
 
 def get_default_lexicon(seed: int | None = None) -> Lexicon:
     """Return the first available lexicon according to configuration priority."""
-
     config = get_config()
     attempts: list[str] = []
     for name in config.lexicon.priority:
@@ -176,6 +182,7 @@ def get_default_lexicon(seed: int | None = None) -> Lexicon:
 
 __all__ = [
     "Lexicon",
+    "LexiconBackend",
     "VectorLexicon",
     "GraphLexicon",
     "WordNetLexicon",

glitchlings/lexicon/_cache.py

@@ -0,0 +1,104 @@
+"""Shared cache helpers for lexicon backends."""
+
+from __future__ import annotations
+
+import json
+from dataclasses import dataclass
+from hashlib import blake2s
+from pathlib import Path
+from typing import Mapping, Sequence
+
+CacheEntries = dict[str, list[str]]
+
+
+@dataclass(frozen=True)
+class CacheSnapshot:
+    """Materialised cache data and its integrity checksum."""
+
+    entries: CacheEntries
+    checksum: str | None = None
+
+
+def _normalise_entries(payload: Mapping[str, Sequence[str]]) -> CacheEntries:
+    """Convert raw cache payloads into canonical mapping form."""
+    entries: CacheEntries = {}
+    for key, values in payload.items():
+        if not isinstance(key, str):
+            raise RuntimeError("Synonym cache keys must be strings.")
+        if not isinstance(values, Sequence):
+            raise RuntimeError("Synonym cache values must be sequences of strings.")
+        entries[key] = [str(value) for value in values]
+    return entries
+
+
+def _canonical_json(entries: Mapping[str, Sequence[str]]) -> str:
+    """Return a deterministic JSON serialisation for ``entries``."""
+    serialisable = {key: list(values) for key, values in sorted(entries.items())}
+    return json.dumps(serialisable, ensure_ascii=False, sort_keys=True, separators=(",", ":"))
+
+
+def compute_checksum(entries: Mapping[str, Sequence[str]]) -> str:
+    """Return a BLAKE2s checksum for ``entries``."""
+    digest = blake2s(_canonical_json(entries).encode("utf8"), digest_size=16)
+    return digest.hexdigest()
+
+
+def load_cache(path: Path) -> CacheSnapshot:
+    """Load a cache from ``path`` and verify its checksum if present."""
+    if not path.exists():
+        return CacheSnapshot(entries={}, checksum=None)
+
+    with path.open("r", encoding="utf8") as handle:
+        payload = json.load(handle)
+
+    checksum: str | None = None
+    entries_payload: Mapping[str, Sequence[str]]
+
+    if isinstance(payload, Mapping) and "__meta__" in payload and "entries" in payload:
+        meta = payload["__meta__"]
+        entries_payload = payload["entries"]  # type: ignore[assignment]
+        if not isinstance(entries_payload, Mapping):
+            raise RuntimeError("Synonym cache entries must be stored as a mapping.")
+        if isinstance(meta, Mapping):
+            raw_checksum = meta.get("checksum")
+            if raw_checksum is not None and not isinstance(raw_checksum, str):
+                raise RuntimeError("Synonym cache checksum must be a string when provided.")
+            checksum = raw_checksum
+        else:
+            raise RuntimeError("Synonym cache metadata must be a mapping.")
+    elif isinstance(payload, Mapping):
+        entries_payload = payload  # legacy format without metadata
+    else:
+        raise RuntimeError("Synonym cache payload must be a mapping of strings to lists.")
+
+    entries = _normalise_entries(entries_payload)
+    if checksum is not None:
+        expected = compute_checksum(entries)
+        if checksum != expected:
+            raise RuntimeError(
+                "Synonym cache checksum mismatch; the cache file appears to be corrupted."
+            )
+
+    return CacheSnapshot(entries=entries, checksum=checksum)
+
+
+def write_cache(path: Path, entries: Mapping[str, Sequence[str]]) -> CacheSnapshot:
+    """Persist ``entries`` to ``path`` with checksum metadata."""
+    serialisable = {key: list(values) for key, values in sorted(entries.items())}
+    checksum = compute_checksum(serialisable)
+    payload = {
+        "__meta__": {
+            "checksum": checksum,
+            "entries": len(serialisable),
+        },
+        "entries": serialisable,
+    }
+    path.parent.mkdir(parents=True, exist_ok=True)
+
+    with path.open("w", encoding="utf8") as handle:
+        json.dump(payload, handle, ensure_ascii=False, indent=2, sort_keys=True)
+
+    return CacheSnapshot(entries=serialisable, checksum=checksum)
+
+
+__all__ = ["CacheEntries", "CacheSnapshot", "compute_checksum", "load_cache", "write_cache"]