Adds constraint-type defaults, partitioner warnings, and model dtype passthrough

CHANGELOG.md

@@ -7,6 +7,24 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [0.9.0] - 2026-07-24
+
+### Added
+
+- `HuggingFaceLanguageModel` and `LanguageModelScorer` accept a `dtype` parameter
+  (default `"auto"`) selecting the precision causal language-model weights load in;
+  `LanguageModelScorer` threads it through to the adapter.
+- `ListPartitioner` warns, once per offending expression or type, when a constraint
+  property expression cannot be evaluated or a constraint type is not enforced during
+  assignment, so a misconfigured constraint no longer fails silently.
+
+### Changed
+
+- `ListConstraint` subclasses now default `constraint_type` to their own discriminator
+  value, so it no longer has to be restated when constructing a constraint.
+- Causal language models load in the checkpoint's own dtype by default rather than being
+  forced to float32, halving the memory used by a half-precision checkpoint.
+
 ## [0.8.0] - 2026-07-13
 
 ### Added
@@ -647,7 +665,8 @@ guards as type-checkers.
 - CI/CD: GitHub Actions for testing, docs, PyPI publishing
 - Read the Docs integration
 
-[Unreleased]: https://github.com/FACTSlab/bead/compare/v0.8.0...HEAD
+[Unreleased]: https://github.com/FACTSlab/bead/compare/v0.9.0...HEAD
+[0.9.0]: https://github.com/FACTSlab/bead/compare/v0.8.0...v0.9.0
 [0.8.0]: https://github.com/FACTSlab/bead/compare/v0.7.0...v0.8.0
 [0.7.0]: https://github.com/FACTSlab/bead/compare/v0.6.0...v0.7.0
 [0.6.0]: https://github.com/FACTSlab/bead/compare/v0.5.0...v0.6.0

bead/__init__.py

@@ -6,6 +6,6 @@
 
 from __future__ import annotations
 
-__version__ = "0.7.0"
+__version__ = "0.9.0"
 __author__ = "Aaron Steven White"
 __email__ = "aaron.white@rochester.edu"

bead/items/adapters/huggingface.py

@@ -59,6 +59,10 @@ class HuggingFaceLanguageModel(HuggingFaceAdapterMixin, ModelAdapter):
         Device to run model on. Falls back to CPU if device unavailable.
     model_version : str
         Version string for cache tracking.
+    dtype : str
+        Torch dtype to load the weights in, such as ``"bfloat16"``. Defaults to
+        ``"auto"``, which keeps the dtype the checkpoint was saved in; loading a
+        half-precision checkpoint as float32 would double its memory.
 
     Examples
     --------
@@ -77,17 +81,21 @@ def __init__(
         cache: ModelOutputCache,
         device: DeviceType = "cpu",
         model_version: str = "unknown",
+        dtype: str = "auto",
     ) -> None:
         super().__init__(model_name, cache, model_version)
         self.device = self._validate_device(device)
+        self.dtype = dtype
         self._model: PreTrainedModel | None = None
         self._tokenizer: PreTrainedTokenizerBase | None = None
 
     def _load_model(self) -> None:
         """Load model and tokenizer lazily on first use."""
         if self._model is None:
             logger.info(f"Loading causal LM: {self.model_name}")
-            self._model = AutoModelForCausalLM.from_pretrained(self.model_name)
+            self._model = AutoModelForCausalLM.from_pretrained(
+                self.model_name, dtype=self.dtype
+            )
             self._model.to(self.device)
             self._model.eval()
 

bead/items/scoring.py

@@ -134,6 +134,9 @@ class LanguageModelScorer(ItemScorer):
         Key in item.rendered_elements to use as text (default: "text").
     model_version : str
         Version string for cache tracking.
+    dtype : str
+        Torch dtype to load the weights in, such as ``"bfloat16"``. Defaults to
+        ``"auto"``, keeping the dtype the checkpoint was saved in.
 
     Examples
     --------
@@ -155,12 +158,14 @@ def __init__(
         device: str = "cpu",
         text_key: str = "text",
         model_version: str = "unknown",
+        dtype: str = "auto",
     ) -> None:
         self.model_name = model_name
         self.cache_dir = Path(cache_dir) if cache_dir else None
         self.device = device
         self.text_key = text_key
         self.model_version = model_version
+        self.dtype = dtype
 
         # lazy loading of model and cache
         self._model: HuggingFaceLanguageModel | None = None
@@ -193,6 +198,7 @@ def model(self) -> HuggingFaceLanguageModel:
                 cache=self._cache,
                 device=self.device,  # type: ignore[arg-type]
                 model_version=self.model_version,
+                dtype=self.dtype,
             )
 
         return self._model

bead/lists/constraints.py

@@ -81,7 +81,7 @@ class UniquenessConstraint(ListConstraint):
         Higher values are weighted more heavily during partitioning.
     """
 
-    constraint_type: typing.Literal["uniqueness"]
+    constraint_type: typing.Literal["uniqueness"] = "uniqueness"
     property_expression: str
     context: dict[str, ContextValue] = dx.field(default_factory=dict)
     allow_null: bool = False
@@ -111,7 +111,7 @@ class ConditionalUniquenessConstraint(ListConstraint):
         Constraint priority.
     """
 
-    constraint_type: typing.Literal["conditional_uniqueness"]
+    constraint_type: typing.Literal["conditional_uniqueness"] = "conditional_uniqueness"
     property_expression: str
     condition_expression: str
     context: dict[str, ContextValue] = dx.field(default_factory=dict)
@@ -140,7 +140,7 @@ class BalanceConstraint(ListConstraint):
         Constraint priority.
     """
 
-    constraint_type: typing.Literal["balance"]
+    constraint_type: typing.Literal["balance"] = "balance"
     property_expression: str
     context: dict[str, ContextValue] = dx.field(default_factory=dict)
     target_counts: dict[str, int] | None = None
@@ -191,7 +191,7 @@ class QuantileConstraint(ListConstraint):
         Constraint priority.
     """
 
-    constraint_type: typing.Literal["quantile"]
+    constraint_type: typing.Literal["quantile"] = "quantile"
     property_expression: str
     context: dict[str, ContextValue] = dx.field(default_factory=dict)
     n_quantiles: int = 5
@@ -231,7 +231,7 @@ class GroupedQuantileConstraint(ListConstraint):
         Constraint priority.
     """
 
-    constraint_type: typing.Literal["grouped_quantile"]
+    constraint_type: typing.Literal["grouped_quantile"] = "grouped_quantile"
     property_expression: str
     group_by_expression: str
     context: dict[str, ContextValue] = dx.field(default_factory=dict)
@@ -468,7 +468,7 @@ class DiversityConstraint(ListConstraint):
         Constraint priority.
     """
 
-    constraint_type: typing.Literal["diversity"]
+    constraint_type: typing.Literal["diversity"] = "diversity"
     property_expression: str
     min_unique_values: int
     context: dict[str, ContextValue] = dx.field(default_factory=dict)
@@ -505,7 +505,7 @@ class SizeConstraint(ListConstraint):
         Constraint priority.
     """
 
-    constraint_type: typing.Literal["size"]
+    constraint_type: typing.Literal["size"] = "size"
     min_size: int | None = None
     max_size: int | None = None
     exact_size: int | None = None
@@ -580,7 +580,7 @@ class OrderingConstraint(ListConstraint):
         Constraint priority (unused for static partitioning).
     """
 
-    constraint_type: typing.Literal["ordering"]
+    constraint_type: typing.Literal["ordering"] = "ordering"
     precedence_pairs: tuple[dx.Embed[OrderingPair], ...] = ()
     no_adjacent_property: str | None = None
     block_by_property: str | None = None
@@ -637,7 +637,7 @@ class BatchCoverageConstraint(BatchConstraint):
         Constraint priority.
     """
 
-    constraint_type: typing.Literal["coverage"]
+    constraint_type: typing.Literal["coverage"] = "coverage"
     property_expression: str
     context: dict[str, ContextValue] = dx.field(default_factory=dict)
     target_values: tuple[str | int | float, ...] | None = None
@@ -673,7 +673,7 @@ class BatchBalanceConstraint(BatchConstraint):
         Constraint priority.
     """
 
-    constraint_type: typing.Literal["balance"]
+    constraint_type: typing.Literal["balance"] = "balance"
     property_expression: str
     target_distribution: dict[str, float]
     context: dict[str, ContextValue] = dx.field(default_factory=dict)
@@ -724,7 +724,7 @@ class BatchDiversityConstraint(BatchConstraint):
         Constraint priority.
     """
 
-    constraint_type: typing.Literal["diversity"]
+    constraint_type: typing.Literal["diversity"] = "diversity"
     property_expression: str
     max_lists_per_value: int
     context: dict[str, ContextValue] = dx.field(default_factory=dict)
@@ -757,7 +757,7 @@ class BatchMinOccurrenceConstraint(BatchConstraint):
         Constraint priority.
     """
 
-    constraint_type: typing.Literal["min_occurrence"]
+    constraint_type: typing.Literal["min_occurrence"] = "min_occurrence"
     property_expression: str
     min_occurrences: int
     context: dict[str, ContextValue] = dx.field(default_factory=dict)

bead/lists/partitioner.py

@@ -7,6 +7,7 @@
 
 from __future__ import annotations
 
+import logging
 from collections import Counter, defaultdict
 from collections.abc import Callable, Hashable
 from typing import Any
@@ -43,6 +44,8 @@
 # without dict-invariance noise.
 type BalanceMetrics = dict[str, "MetadataValue"]
 
+logger = logging.getLogger(__name__)
+
 
 class ListPartitioner:
     """Partitions items into balanced experimental lists.
@@ -80,6 +83,8 @@ def __init__(self, random_seed: int | None = None) -> None:
         self.random_seed = random_seed
         self._rng = np.random.default_rng(random_seed)
         self.dsl_evaluator = DSLEvaluator()
+        self._unresolved_expressions: set[str] = set()
+        self._unenforced_types: set[str] = set()
 
     def partition(
         self,
@@ -448,6 +453,8 @@ def _count_violations(
             elif isinstance(constraint, SizeConstraint):
                 if not self._check_size(exp_list, constraint):
                     is_violated = True
+            else:
+                self._warn_unenforced(str(constraint.constraint_type))
 
             if is_violated:
                 priority = constraint.priority
@@ -574,6 +581,50 @@ def _check_size(self, exp_list: ExperimentList, constraint: SizeConstraint) -> b
 
         return True
 
+    def _warn_unresolved(self, property_expression: str, error: Exception) -> None:
+        """Report a property expression that cannot be evaluated.
+
+        Scoring skips items whose property cannot be read. Without a warning a
+        misconfigured constraint scores zero on every iteration and silently has
+        no effect, so each expression is reported once.
+
+        Parameters
+        ----------
+        property_expression : str
+            The expression that failed to evaluate.
+        error : Exception
+            The failure raised while evaluating it.
+        """
+        if property_expression in self._unresolved_expressions:
+            return
+        self._unresolved_expressions.add(property_expression)
+        logger.warning(
+            "Constraint property expression %r could not be evaluated (%s). "
+            "The constraint cannot be satisfied and is being ignored.",
+            property_expression,
+            error,
+        )
+
+    def _warn_unenforced(self, constraint_type: str) -> None:
+        """Report a constraint type that assignment cannot enforce.
+
+        Unhandled types are stored on the list and serialised, so without a
+        warning a configured constraint appears active while having no effect.
+
+        Parameters
+        ----------
+        constraint_type : str
+            Discriminator of the constraint that is not enforced.
+        """
+        if constraint_type in self._unenforced_types:
+            return
+        self._unenforced_types.add(constraint_type)
+        logger.warning(
+            "List constraint type %r is not enforced during assignment. It is "
+            "recorded on the list but does not influence partitioning.",
+            constraint_type,
+        )
+
     def _extract_property_value(
         self,
         item_id: UUID,
@@ -1034,7 +1085,8 @@ def _compute_batch_coverage_score(
                         metadata,
                     )
                     observed_values.add(value)
-                except Exception:
+                except Exception as error:
+                    self._warn_unresolved(constraint.property_expression, error)
                     continue
 
         # Compute coverage
@@ -1085,7 +1137,8 @@ def _compute_batch_balance_score(
                     )
                     counts[value] += 1
                     total += 1
-                except Exception:
+                except Exception as error:
+                    self._warn_unresolved(constraint.property_expression, error)
                     continue
 
         if total == 0:
@@ -1140,7 +1193,8 @@ def _compute_batch_diversity_score(
                         metadata,
                     )
                     value_to_lists[value].add(list_idx)
-                except Exception:
+                except Exception as error:
+                    self._warn_unresolved(constraint.property_expression, error)
                     continue
 
         if not value_to_lists:
@@ -1193,7 +1247,8 @@ def _compute_batch_min_occurrence_score(
                         metadata,
                     )
                     counts[value] += 1
-                except Exception:
+                except Exception as error:
+                    self._warn_unresolved(constraint.property_expression, error)
                     continue
 
         if not counts:

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "bead"
-version = "0.8.0"
+version = "0.9.0"
 description = "Lexicon and Template Collection Construction Pipeline for Acceptability and Inference Judgment Data"
 authors = [{name = "Aaron Steven White", email = "aaron.white@rochester.edu"}]
 readme = "README.md"