Fix RUNTIME_ERROR: handle {name, category} dicts in models.benchmarks

data.py

@@ -1,8 +1,10 @@
 """Load and flatten the SaylorTwift/llm-benchmark-usage dataset for exploration.
 
 The dataset has two configs:
-- `models`: one row per model, `benchmarks` is a flat list of canonicalized names
-  (no per-model category — that free-text field was dropped when this was flattened).
+- `models`: one row per model, `benchmarks` is a list of canonicalized names — either
+  plain strings or `{name, category}` dicts (the `category` there is a free-text,
+  per-source label and is not used here; the controlled-vocabulary category used
+  for aggregation always comes from the `benchmarks` reference table below).
 - `benchmarks`: reference metadata for a subset of those names — `categories`
   (controlled vocabulary, can be more than one, e.g. a vision-math benchmark is
   both `math` and `perception`), `modality`, `language`, `description`,
@@ -60,7 +62,8 @@ def load_data():
     rows = []
     for m in models_raw:
         meta = model_meta[m["model_id"]]
-        for name in m["benchmarks"]:
+        for b in m["benchmarks"]:
+            name = b["name"] if isinstance(b, dict) else b
             info = bench_meta.get(name)
             categories = info["categories"] if info and info["categories"] else ["uncategorized"]
             for cat in categories: