ignore padding tokens in surprisal aggregation

syntaxgym.py

@@ -285,6 +285,12 @@ class SyntaxGymMetric(datasets.Metric):
                 # token = i_tokens[t_cursor]
                 token_char_start, token_char_end = i_offsets[t_cursor]
 
+                if token_char_start == token_char_end == 0:
+                    # This is a padding token. Skip.
+                    # TODO what about BOS/EOS? some models incorporate them
+                    t_cursor += 1
+                    continue
+
                 region_start = region_edges[r_cursor]
                 region_end = region_edges[r_cursor + 1] \
                     if r_cursor + 1 < len(region_edges) else max_long