metric.py

from datasets import Metric

class CUEBenchMetric(Metric):
    def _info(self):
        return {
            "description": "F1, Precision, and Recall for multi-label set prediction in CUEBench",
            "inputs_description": "List of predicted and reference class sets",
            "citation": "",
        }

    def _MeanReciprocalRank(self, predicted, target):
        if not predicted or not target:
            return 0
        predicted = [str(p).lower() for p in predicted]
        target = [str(t).lower() for t in target]
        for i, p in enumerate(predicted):
            if p in target:
                return 1 / (i + 1)
        return 0

    def _Hits_at_K(self, predicted, target, k):
        if not predicted or not target:
            return 0
        predicted = [str(p).lower() for p in predicted]
        target = [str(t).lower() for t in target]
        return sum(1 for p in predicted[:k] if p in target)

    def _coverage(self, _pd_Res, _eGold, _scores=None):
        """
        Evaluate predictions (_pd_Res) against gold labels (_eGold).
        Optionally, pass _scores (same length as _pd_Res) if you want to track prediction scores.
        
        Returns:
            res: [cov@len(_eGold), cov@1, cov@3, cov@5, rank_first_gold]
            l_gold_pred: (_eGold, (top_predicted_labels, top_scores)) at len(_eGold)
        """
        res = {}
        l_gold_pred = ()
        
        if not _pd_Res or not _eGold:
            for k in [1, 3, 5, 10]:
                res[k] = 0
                
            # res.append(rank_first_gold)
            return res, l_gold_pred

        all_labels = _pd_Res

        # Check if there's any overlap between predicted and gold labels
        if set(_eGold) & set(all_labels):
            # Find the 1-based rank of the first correct prediction
            rank_first_gold = min([r + 1 for r, l in enumerate(all_labels) if l in _eGold])

            for k in [1, 3, 5, 10]:
                top_k_labels = all_labels[:k]
                overlap = set(top_k_labels) & set(_eGold)
                cov_k = len(overlap) / k
                res[k] = (cov_k)

                if k >= len(_eGold):
                    top_scores = _scores[:k] if _scores else None
                    l_gold_pred = (_eGold, (top_k_labels, top_scores))

            # res.append(rank_first_gold)
            return res, l_gold_pred
        else:
            for k in [1, 3, 5, 10]:
                res[k] = 0
                
            # res.append(rank_first_gold)
            return res, l_gold_pred


    def _clean(self, strings):
        cleaned = []
        for s in strings:
            # Remove all asterisks and extra whitespace first
            s = s.replace('*', '').strip()
            
            # Remove surrounding quotes if they match (both single or both double)
            if (s.startswith("'") and s.endswith("'")) or (s.startswith('"') and s.endswith('"')):
                s = s[1:-1]
                
            # Remove square brackets
            s = s.replace('[', '').replace(']', '')
            
            # Handle colon case - take the part after last colon and clean it
            if ':' in s:
                s = s.split(':')[-1]
            
            # Final cleanup - remove any remaining special chars and whitespace
            s = s.strip(' _\\"\'')
            
            cleaned.append(s)
        return cleaned

    def _compute(self, outputs):
        for i in range(len(outputs)):
            outputs[i]['predicted_classes'] = self._clean(outputs[i]['predicted_classes'])
            
        average_mrr = 0
        for i in outputs:
            average_mrr += self._MeanReciprocalRank(i['predicted_classes'], i['target_classes'])

        average_mrr = average_mrr / len(outputs)

        hits_at_1 = 0
        hits_at_3 = 0
        hits_at_5 = 0
        hits_at_10 = 0

        for i in outputs:
            hits_at_1 += 1 if self._Hits_at_K(i['predicted_classes'], i['target_classes'], 1) > 0 else 0
            hits_at_3 += 1 if self._Hits_at_K(i['predicted_classes'], i['target_classes'], 3) > 0 else 0
            hits_at_5 += 1 if self._Hits_at_K(i['predicted_classes'], i['target_classes'], 5) > 0 else 0
            hits_at_10 += 1 if self._Hits_at_K(i['predicted_classes'], i['target_classes'], 10) > 0 else 0

        hits_at_1 = hits_at_1 / len(outputs)
        hits_at_3 = hits_at_3 / len(outputs)
        hits_at_5 = hits_at_5 / len(outputs)
        hits_at_10 = hits_at_10 / len(outputs)

        cov_1 = 0
        cov_3 = 0
        cov_5 = 0
        cov_10 = 0

        for i in outputs:
            res, l_gold_pred = self._coverage(i['predicted_classes'], i['target_classes'])
            cov_1 += res[1]
            cov_3 += res[3]
            cov_5 += res[5]
            cov_10 += res[10]

        cov_1 = cov_1 / len(outputs)
        cov_3 = cov_3 / len(outputs)
        cov_5 = cov_5 / len(outputs)
        cov_10 = cov_10 / len(outputs)
        
        return {
            "average_mrr": average_mrr,
            "hits_at_1" : hits_at_1,
            "hits_at_3" : hits_at_3,
            "hits_at_5" : hits_at_5,
            "hits_at_10" : hits_at_10,
            "coverage_at_1" : cov_1,
            "coverage_at_3" : cov_3,
            "coverage_at_5" : cov_5,
            "coverage_at_10" : cov_10,
        }