In py/games/connect4/spec.py, we have Connect4Spec, which has a training_params member specifying default training parameters.
When you launch run_local.py or run_loop_controller.py, those values are used as the defaults if you pass -g c4. Please see the call to TrainingParams.add_args() inside of run_loop_controller.py to see how this is done.
We should use this exact same mechanism to make sure that the -i: 100 and -n: 4 specified in Connect4Spec are used in benchmark/eval runs.
As it stands, benchmark_tag_local.py is multiple times slower than it should be because its failing to use -n 4.