Feat/bias evaluator WinoBias (Gender)

openverifiablellm/eval/__init__.py

@@ -0,0 +1,7 @@
+from .bias import WinoBiasEvaluator
+from .perplexity import PerplexityEvaluator
+
+__all__ = [
+    "WinoBiasEvaluator",
+    "PerplexityEvaluator",
+]

openverifiablellm/eval/base.py

@@ -0,0 +1,24 @@
+from abc import ABC, abstractmethod
+
+
+class BaseEvaluator(ABC):
+    """Abstract base class for all dataset evaluators."""
+
+    @abstractmethod
+    def evaluate(self, model, tokenizer) -> dict:
+        """
+        Evaluate a language model using the given tokenizer.
+
+        Parameters
+        ----------
+        model : callable
+            Callable accepting a sequence of token IDs and returning a
+            2-D sequence of logits with shape ``(len(input_ids), vocab_size)``.
+        tokenizer : object
+            Object with an ``encode(text: str) -> list[int]`` method.
+
+        Returns
+        -------
+        dict
+            Benchmark-specific evaluation results.
+        """

openverifiablellm/eval/bias/__init__.py

@@ -0,0 +1,5 @@
+from .wino_bias import WinoBiasEvaluator
+
+__all__ = [
+    "WinoBiasEvaluator",
+]

openverifiablellm/eval/bias/wino_bias.py

@@ -0,0 +1,91 @@
+"""
+openverifiablellm/eval/bias/wino_bias.py
+
+Gender-bias evaluator using the WinoBias benchmark.
+"""
+
+from typing import Optional
+
+from ..base import BaseEvaluator
+from ..perplexity import PerplexityEvaluator
+
+
+class WinoBiasEvaluator(BaseEvaluator):
+    """
+    Evaluates gender bias in a language model using the WinoBias benchmark.
+
+    For each sentence pair (pro-stereotype / anti-stereotype) the model's
+    perplexity is computed via the same sliding-window method used by
+    :class:`PerplexityEvaluator`.  A lower ``bias_score`` indicates a less
+    biased model.
+
+    Parameters
+    ----------
+    n_samples : int or None
+        Maximum number of sentences to load from each WinoBias split.
+        ``None`` evaluates the full dataset.  Default ``None``.
+    """
+
+    def __init__(self, n_samples: Optional[int] = None):
+        self.n_samples = n_samples
+
+    def evaluate(self, model, tokenizer) -> dict:
+        """
+        Compute stereotype and anti-stereotype perplexity scores.
+
+        Loads ``type1_pro`` (pro-stereotype) and ``type1_anti``
+        (anti-stereotype) splits of WinoBias and measures how much more
+        easily the model predicts gender-stereotypical sentences than
+        counter-stereotypical ones.
+
+        Parameters
+        ----------
+        model : callable
+            ``model(input_ids) -> 2-D sequence`` of shape
+            ``(len(input_ids), vocab_size)``, as described in
+            :meth:`PerplexityEvaluator.compute_sentence_perplexity`.
+        tokenizer : object
+            Object with ``encode(text: str) -> list[int]``.
+
+        Returns
+        -------
+        dict
+            A dictionary with the following keys:
+
+            * **stereotype_score** (*float*) — mean perplexity on
+              pro-stereotype sentences.
+            * **anti_stereotype_score** (*float*) — mean perplexity on
+              anti-stereotype sentences.
+            * **bias_score** (*float*) —
+              ``abs(stereotype_score - anti_stereotype_score)``;
+              lower means less biased.
+        """
+        import datasets as hf_datasets  # deferred; runtime dep
+
+        pro_ds = hf_datasets.load_dataset("wino_bias", "type1_pro", split="test")
+        anti_ds = hf_datasets.load_dataset("wino_bias", "type1_anti", split="test")
+
+        def _score_split(dataset) -> float:
+            scores = []
+            for i, row in enumerate(dataset):
+                if self.n_samples is not None and i >= self.n_samples:
+                    break
+                tokens = row.get("tokens", [])
+                text = " ".join(tokens) if isinstance(tokens, list) else str(tokens)
+                if not text.strip():
+                    continue
+                token_ids = tokenizer.encode(text)
+                scores.append(
+                    PerplexityEvaluator.compute_sentence_perplexity(model, token_ids)
+                )
+            return float(sum(scores) / len(scores)) if scores else float("inf")
+
+        stereotype_score = _score_split(pro_ds)
+        anti_stereotype_score = _score_split(anti_ds)
+        bias_score = abs(stereotype_score - anti_stereotype_score)
+
+        return {
+            "stereotype_score": stereotype_score,
+            "anti_stereotype_score": anti_stereotype_score,
+            "bias_score": bias_score,
+        }

openverifiablellm/eval/perplexity.py

@@ -0,0 +1,151 @@
+"""
+openverifiablellm/eval/perplexity.py
+
+Perplexity evaluator for language models.
+"""
+
+import math
+from typing import List, Optional
+
+from .base import BaseEvaluator
+
+
+class PerplexityEvaluator(BaseEvaluator):
+    """
+    Evaluates language-model perplexity on a HuggingFace benchmark dataset.
+
+    Perplexity is computed with a teacher-forced sliding-window approach:
+    for each token position *i* the model receives tokens ``[0 .. i-1]``
+    and the negative log-probability of token ``[i]`` is accumulated.
+    The final perplexity is ``exp(mean_NLL)``.
+
+    Parameters
+    ----------
+    benchmark : str
+        HuggingFace dataset identifier.  Default ``"wikitext"``.
+    n_samples : int or None
+        Maximum number of non-empty samples to evaluate.  ``None`` means
+        evaluate the whole dataset.  Default ``50``.
+    stride : int
+        Window stride used when the sequence exceeds the model's context
+        window.  Default ``512``.
+    """
+
+    def __init__(
+        self,
+        benchmark: str = "wikitext",
+        n_samples: Optional[int] = 50,
+        stride: int = 512,
+    ):
+        self.benchmark = benchmark
+        self.n_samples = n_samples
+        self.stride = stride
+
+    # ------------------------------------------------------------------
+    # Mock helpers
+    # ------------------------------------------------------------------
+
+    @staticmethod
+    def uniform_model(vocab_size: int = 1000):
+        """
+        Return a mock model that produces uniform (all-zero) logits.
+
+        Useful for unit testing: because all logits are equal, the
+        log-softmax is ``-log(vocab_size)`` at every position, giving a
+        predictable perplexity of exactly ``vocab_size``.
+
+        Parameters
+        ----------
+        vocab_size : int
+            Vocabulary size of the mock model.  Default ``1000``.
+
+        Returns
+        -------
+        callable
+            ``model(input_ids) -> list[list[float]]`` of shape
+            ``(len(input_ids), vocab_size)``.
+        """
+
+        def _model(input_ids):
+            return [[0.0] * vocab_size for _ in input_ids]
+
+        return _model
+
+    # ------------------------------------------------------------------
+    # Core computation
+    # ------------------------------------------------------------------
+
+    @staticmethod
+    def compute_sentence_perplexity(model, token_ids: List[int]) -> float:
+        """
+        Compute the perplexity of *token_ids* under *model*.
+
+        Parameters
+        ----------
+        model : callable
+            ``model(input_ids) -> 2-D sequence`` of shape
+            ``(len(input_ids), vocab_size)``.
+        token_ids : list[int]
+            Tokenised sentence.
+
+        Returns
+        -------
+        float
+            Perplexity (≥ 1).  Returns ``float("inf")`` for sequences
+            shorter than 2 tokens.
+        """
+        if len(token_ids) < 2:
+            return float("inf")
+
+        inputs = token_ids[:-1]
+        targets = token_ids[1:]
+
+        logits_batch = model(inputs)  # shape: (n-1, vocab_size)
+
+        nll_sum = 0.0
+        for logits, target in zip(logits_batch, targets):
+            # numerically-stable log-softmax
+            max_l = max(logits)
+            exp_shifted = [math.exp(v - max_l) for v in logits]
+            log_sum = math.log(sum(exp_shifted))
+            log_prob_target = (logits[target] - max_l) - log_sum
+            nll_sum -= log_prob_target
+
+        return math.exp(nll_sum / len(targets))
+
+    # ------------------------------------------------------------------
+    # BaseEvaluator interface
+    # ------------------------------------------------------------------
+
+    def evaluate(self, model, tokenizer) -> dict:
+        """
+        Compute mean perplexity on *self.benchmark*.
+
+        Parameters
+        ----------
+        model : callable
+            Callable as described in :meth:`compute_sentence_perplexity`.
+        tokenizer : object
+            Object with ``encode(text: str) -> list[int]``.
+
+        Returns
+        -------
+        dict
+            ``{"perplexity": float}`` — mean perplexity across evaluated
+            sentences.
+        """
+        import datasets as hf_datasets  # deferred; runtime dep
+
+        ds = hf_datasets.load_dataset(self.benchmark, split="test", streaming=True)
+        scores = []
+        for i, row in enumerate(ds):
+            if self.n_samples is not None and i >= self.n_samples:
+                break
+            text = row.get("text", "")
+            if not text.strip():
+                continue
+            token_ids = tokenizer.encode(text)
+            scores.append(self.compute_sentence_perplexity(model, token_ids))
+
+        mean_ppl = float(sum(scores) / len(scores)) if scores else float("inf")
+        return {"perplexity": mean_ppl}

pyproject.toml

@@ -12,14 +12,25 @@ authors = [
 requires-python = ">=3.9"
 
 dependencies = [
+    "datasets",
     "defusedxml",
     "sentencepiece",
     "tokenizers==0.15.2"
 ]
 
+# Intentionally duplicated from [dependency-groups] below.
+# pip uses this section; uv/PEP 735 uses [dependency-groups]. Keep both in sync.
+[project.optional-dependencies]
+dev = [
+    "pytest>=7.0",
+    "ruff>=0.15.4",
+]
+
 [tool.setuptools.packages.find]
 include = ["openverifiablellm*"]
 
+# Intentionally duplicated from [project.optional-dependencies] above.
+# uv/PEP 735 uses this section; pip uses [project.optional-dependencies]. Keep both in sync.
 [dependency-groups]
 dev = [
     "pytest>=7.0",