Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@openverifiablellm/eval/bias/wino_bias.py`:
- Around line 83-86: The computation of bias_score uses stereotype_score and
anti_stereotype_score from _score_split, and when both return inf the
subtraction yields NaN; update the logic after computing stereotype_score and
anti_stereotype_score to detect when both are infinite (e.g., via math.isfinite
or math.isinf) and in that case set bias_score to a stable sentinel like
float("inf") instead of performing the subtraction; otherwise keep the existing
abs(stereotype_score - anti_stereotype_score) behavior so normal finite values
are unchanged.

In `@openverifiablellm/eval/perplexity.py`:
- Around line 29-43: The stride parameter is stored in __init__ but never used;
update the evaluator's sequence-processing logic (e.g., the method that prepares
token windows or computes perplexity—look for methods like evaluate, score, or
compute_perplexity) to respect self.stride when a text exceeds the model context
length by implementing sliding windows with overlap equal to self.stride (or
using stride to advance the window), ensure tokenization/truncation uses these
windows rather than a single chop, and keep the docstring and self.stride
assignment in sync; also add tests or a small example in the evaluate path to
verify long sequences are segmented with the configured stride.
- Around line 141-149: The loop in perplexity evaluation (in
openverifiablellm.eval.perplexity) checks self.n_samples before filtering empty
texts so blank rows consume the quota; change the logic to count only actual
evaluated samples by moving the n_samples check after the empty-text check (or
maintain an explicit evaluated counter that increments only when you call
compute_sentence_perplexity and append to scores). In practice, inside the
for-loop (where tokenizer.encode, compute_sentence_perplexity, and scores.append
are used), skip empty texts first, then check if the evaluated sample count (use
len(scores) or a separate counter) has reached self.n_samples and break,
ensuring compute_sentence_perplexity is only called for non-empty rows and the
n_samples limit counts evaluated samples.
- Around line 103-114: The code currently zips logits_batch and targets which
silently truncates when model output length is wrong; update the scoring logic
in perplexity.py (around the use of logits_batch, model, targets and the loop
computing nll_sum) to first validate that len(logits_batch) == len(targets) and
raise a clear exception (e.g., ValueError) showing both lengths if they differ,
and optionally also validate that each logits has length > max(target) or
matches expected vocab_size before computing log-probs; only proceed to the
existing loop and return math.exp(nll_sum / len(targets)) after these checks.

In `@pyproject.toml`:
- Around line 14-19: The dependencies list in pyproject.toml is missing core LLM
packages; add "numpy" and "torch" to the existing dependencies array (alongside
"datasets", "defusedxml", "sentencepiece", "tokenizers==0.15.2") so they are
installed as required dependencies; update the dependencies section to include
"numpy" and "torch" entries and ensure versions are specified if necessary for
compatibility.

In `@tests/test_eval.py`:
- Around line 48-52: The test helper _load currently returns anti_data for any
unexpected inputs which can mask integration bugs; update the mock _load
function to strictly validate its inputs (check the incoming name, config, and
split against the expected values) and raise an explicit exception (e.g.,
ValueError or AssertionError) when an unexpected name/config/split is passed
instead of defaulting to anti_data; adjust references to pro_data/anti_data
accordingly and ensure tests that expect fallback behavior are updated,
mentioning the mock function name _load (and any callers like evaluate) so
reviewers can locate the change.

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@openverifiablellm/eval/bias/wino_bias.py`:
- Around line 78-81: The current averaging in the block that calls
PerplexityEvaluator.compute_sentence_perplexity can produce float("inf") if any
sentence returns infinity; update the logic in the loop/return that collects
scores (where scores is appended using
PerplexityEvaluator.compute_sentence_perplexity(model, token_ids)) to filter out
math.isfinite or not-infinite values before computing the mean, and if no finite
scores remain return float("inf") (or keep the original fallback) so a single
infinite sentence doesn't make the whole split score infinite.

---

Duplicate comments:
In `@openverifiablellm/eval/bias/wino_bias.py`:
- Line 83: The import math is inside the evaluate() function after the nested
_score_split definition; move the import to the module level (with the other
imports at the top of the file) and remove the in-function "import math"
statement so evaluate() and _score_split use the top-level import instead.