def print_samples()

in lm_human_preferences/train_policy.py [0:0]

• 13 lines of code
• 4 McCabe index (conditional complexity)

    def print_samples(self, queries, responses, scores, logprobs, ref_logprobs):
        if self.comm.Get_rank() != 0:
            return
        if tf.train.get_global_step().eval() % self.hparams.run.log_interval != 0:
            return

        encoder = self.policy.encoder

        # Log samples
        for i in range(min(3, len(queries))):
            sample_kl = np.sum(logprobs[i] - ref_logprobs[i])
            print(encoder.decode(queries[i][:self.hparams.task.query_length]).replace("\n", "⏎"))
            print(encoder.decode(responses[i]).replace("\n", "⏎"))
            print(f"  score = {scores[i]:+.2f}")
            print(f"  kl = {sample_kl:+.2f}")
            print(f"  total = {scores[i] - self.hparams.rewards.kl_coef * sample_kl:+.2f}")