Method process_results

lm-eval-harness/lm_eval/api/task.py:1176–1355

(self, doc, results)

        )

    def process_results(self, doc, results):
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)

        result_dict = {}
        use_metric = list(self._metric_fn_list.keys())
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
            (loglikelihood,) = results
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
            return {
                **(
                    {"word_perplexity": (loglikelihood, _words)}
                    if "word_perplexity" in use_metric
                    else {}
                ),
                **(
                    {"byte_perplexity": (loglikelihood, _bytes)}
                    if "byte_perplexity" in use_metric
                    else {}
                ),
                **(
                    {"bits_per_byte": (loglikelihood, _bytes)}
                    if "bits_per_byte" in use_metric
                    else {}
                ),
            }
        elif self.OUTPUT_TYPE == "multiple_choice":
            lls, is_greedy = zip(*results)

            # retrieve choices in List[str] form, to compute choice lengths, etc.
            choices = self.doc_to_choice(doc)
            completion_len = np.array([float(len(i)) for i in choices])

            if (
                2 * len(choices) == len(lls)
                and "acc_mutual_info" in self._metric_fn_list.keys()
            ):
                # then we are doing mutual info.
                # this stores the "dryrun" / unconditional answer loglikelihoods
                lls_unconditional = lls[1::2]
                assert len(lls_unconditional) == len(choices)
                # and this stores our "regular" conditional loglikelihoods
                lls = lls[::2]

            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)

            if self.multiple_input:
                gold = self.doc_to_text(doc)
            else: