Model training:
python run_experiment.py --run_dir runs/repro \
--learner SupervisedLearner \
--train_file data/selection_repro_train.jsons \
--validation_file data/selection_repro_val.jsons \
--eval_file data/selection_repro_test.jsons \
--metrics response_log_likelihood_bits \
response_perplexity \
response_token_perplexity_micro \
selection_log_likelihood_bits \
selection_accuracy \
selection_perplexity
Self-play dialogue simulation:
python repl.py --run_dir runs/sim_repro \
--agent_a FBReproAgent --agent_b FBReproAgent \
--load_a runs/repro/model.pkl --load_b runs/repro/model.pkl \
--contexts data/selfplay.txt