olmOCRBench: `evaluate()` returns `None`, so the score never reaches `status.json`

Summary

OlmOCRBench.evaluate() computes its results and writes them to disk, but does
not return them. run.py uses the return value to record the score, so the run
ends with

"skip_reason": "evaluate_returned_none"

while olmOCRBench_eval_summary.csv sits next to the prediction file
containing the correct numbers. The benchmark has effectively run successfully
and reported nothing.

Every other dataset's evaluate() returns a DataFrame or dict; this one is
the exception.

Where

vlmeval/dataset/olmOCRBench/olmocrbench.py:

def evaluate(self, eval_file, **judge_kwargs):
    try:
        from .evaluator import evaluator
    except ImportError as e:
        ...
        raise e
    tsv_path = self.data_path
    evaluator(tsv_path, eval_file)      # <- no return

evaluator() writes the summary as a side effect —
vlmeval/dataset/olmOCRBench/evaluator.py:449:

csv_path = os.path.join(os.path.dirname(eval_file), "olmOCRBench_eval_summary.csv")
rows = [["type", "score"]]
rows.append(["overall", f"{new_overall_score * 100:.1f}"])
for jsonl_file, results in sorted(jsonl_results.items()):
    if results["total"] > 0:
        rows.append([jsonl_file, f"{pass_rate:.1f}"])

Impact

• status.json records no score for olmOCRBench, so scripts/summarize.py
  reports nothing for it. Its main() returns silently when summary_rows is
  empty, so this presents as an absent dataset rather than an error.
• Anything that aggregates across datasets treats a completed, correct run as a
  missing one.
• The failure is easy to misread as an evaluation crash, because the
  skip_reason says the evaluation was skipped when in fact it completed.

Suggested fix

Return the summary that evaluator() already produces. Minimal version, which
re-reads the file the evaluator just wrote:

     def evaluate(self, eval_file, **judge_kwargs):
         try:
             from .evaluator import evaluator
         except ImportError as e:
             ...
             raise e
         tsv_path = self.data_path
-        evaluator(tsv_path, eval_file)
+        evaluator(tsv_path, eval_file)
+        csv_path = osp.join(osp.dirname(eval_file), 'olmOCRBench_eval_summary.csv')
+        if osp.exists(csv_path):
+            df = pd.read_csv(csv_path)
+            # {'overall': 63.4, 'tables': 55.0, ...}
+            return dict(zip(df['type'], df['score']))
+        return None

Cleaner would be for evaluator() to return the rows it writes, and for
evaluate() to pass them through — that avoids the write/read round trip
entirely.

A related naming point

The summary is written as a fixed, unprefixed filename:

olmOCRBench_eval_summary.csv

rather than the <model>_<dataset>_acc.csv convention the other datasets
follow. That means:

• tooling that locates score files by the usual <model>_<dataset>* pattern
  cannot find it
• two models evaluated into the same directory would overwrite each other's
  summary

Renaming it to {eval_file_stem}_acc.csv — or using
get_intermediate_file_path(eval_file, '_acc'), as other datasets do — would
make it consistent and collision-free. Happy to include that in the same PR if
it is wanted, or leave it out to keep the change minimal.

Environment

• VLMEvalKit: main
• Reproduced with a local model through the --mode eval path; independent of
  model and backend, since it is purely the missing return value.

Happy to open a PR.

Start in vlmeval/dataset/olmOCRBench/olmocrbench.py and trace how run.py records evaluate() results. Read evaluator.py around line 449 and scripts/summarize.py, then run the --mode eval path and inspect status.json. Done means olmOCRBench reports its computed summary instead of evaluate_returned_none; filename consistency is an optional follow-up.