def respond_op()

in lm_human_preferences/policy.py [0:0]

• 16 lines of code
• 1 McCabe index (conditional complexity)

    def respond_op(self, queries, length):
        contexts = self.embed_queries(queries)
        context_length = tf.shape(contexts)[1]
        result = sample.sample_sequence(
            step=self.step_core,
            context=contexts,
            length=length,
            model_hparams=self.model_hparams,
            temperature=self.temperature,
            extra_outputs={'values':tf.float32},
        )
        return dict(
            responses=result['tokens'][:, context_length:],
            logprobs=result['logprobs'],
            values=result['values'],
        )