.

code_eval.py

@@ -188,8 +188,8 @@ class CodeEval(evaluate.Metric):
                     body = Template(pred_template).safe_substitute(prediction=p)
                     for r in ref:
                         assert isinstance(r, str)
-                        test = Template(ref_template).safe_substitute(ref_key=r)
-                        test = Template(test).safe_substitute(reference=p)
+                        test = Template(ref_template).safe_substitute(reference=r)
+                        test = Template(test).safe_substitute(prediction=p)
 
                         test_program = body + "\n" + test
                         args = (test_program, timeout, tid, pid)