OpenDataArena
| {{ $t('lb_rank') }} | {{ $t('lb_dataset') }} | {{ $t('lb_year') }} | {{ $t('lb_size') }} | {{ $t('lb_avg_score') }} |
{{ $t('lb_efficiency') }}
{{ $t('lb_efficiency_tip') }}
|
{{ $t('all_general') }} | {{ $t('all_math') }} | {{ $t('all_code') }} | {{ $t('all_reasoning') }} |
|---|---|---|---|---|---|---|---|---|---|
|
{{ getRank(dataset, filteredDataForRanking, (sortColumn
=== 'year' || sortColumn === 'name' || sortColumn ===
'size') ? 'overall_avg' : sortColumn, false, null,
improvementType) }}
{{ dataset.domain === 'base' ?
$t('lb_tooltip_baseline_base') :
$t('lb_tooltip_baseline_instruct') }}
|
{{ dataset.affiliation }}
|
{{ dataset.domain === 'base' || dataset.domain === 'instruct' ? '-' : dataset.year }} | {{ dataset.domain === 'base' || dataset.domain === 'instruct' ? '-' : dataset.size }} |
{{ formatScoreWithImprovement(dataset.overall_avg,
dataset.improvement &&
dataset.improvement.overall_avg, dataset,
improvementType).score }}
{{ formatScoreWithImprovement(dataset.overall_avg,
dataset.improvement &&
dataset.improvement.overall_avg, dataset,
improvementType).diffText }}
|
{{
formatEfficiencyScore(dataset.overall_efficiency)
}}
|
{{ formatScoreWithImprovement(dataset.general_avg,
dataset.improvement &&
dataset.improvement.general_avg, dataset,
improvementType).score }}
{{ formatScoreWithImprovement(dataset.general_avg,
dataset.improvement &&
dataset.improvement.general_avg, dataset,
improvementType).diffText }}
|
{{ formatScoreWithImprovement(dataset.math_avg,
dataset.improvement && dataset.improvement.math_avg,
dataset, improvementType).score }}
{{ formatScoreWithImprovement(dataset.math_avg,
dataset.improvement && dataset.improvement.math_avg,
dataset, improvementType).diffText }}
|
{{ formatScoreWithImprovement(dataset.code_avg,
dataset.improvement && dataset.improvement.code_avg,
dataset, improvementType).score }}
{{ formatScoreWithImprovement(dataset.code_avg,
dataset.improvement && dataset.improvement.code_avg,
dataset, improvementType).diffText }}
|
{{
formatScoreWithImprovement(dataset.reasoning_avg,
dataset.improvement &&
dataset.improvement.reasoning_avg, dataset,
improvementType).score }}
{{ formatScoreWithImprovement(dataset.reasoning_avg,
dataset.improvement &&
dataset.improvement.reasoning_avg, dataset,
improvementType).diffText }}
|
{{ $t('all_' + selectedType.toLowerCase()) !== ('all_' + selectedType.toLowerCase()) ? $t('all_' + selectedType.toLowerCase()) : getTypeDisplayName(selectedType) }}{{ $t('all_leaderboard') }}
| {{ $t('lb_rank') }} | {{ $t('lb_dataset') }} | {{ $t('lb_year') }} | {{ $t('lb_size') }} | {{ $t('lb_avg_score') }} |
{{ $t('lb_efficiency') }}
{{ $t('lb_efficiency_tip') }}
|
{{ header.displayName }}
{{ header.metricDisplayName }}
|
|---|---|---|---|---|---|---|
|
{{ getRank(dataset, detailedFilteredDataForRanking,
(detailedSortColumn === 'year' || detailedSortColumn ===
'name' || detailedSortColumn === 'size') ? 'average' :
(detailedSortColumn || 'average'), true, selectedType,
improvementType) }}
{{ dataset.domain === 'base' ?
$t('lb_tooltip_baseline_base') :
$t('lb_tooltip_baseline_instruct') }}
|
{{ dataset.affiliation }}
|
{{ dataset.domain === 'base' || dataset.domain === 'instruct' ? '-' : dataset.year }} | {{ dataset.domain === 'base' || dataset.domain === 'instruct' ? '-' : dataset.size }} |
{{ getTypeAverage(dataset, selectedType,
improvementType).score }}
{{ getTypeAverage(dataset, selectedType,
improvementType).diffText }}
|
{{ formatEfficiencyScore(getTypeEfficiency(dataset,
selectedType)) }}
|
{{ getTaskScore(dataset, selectedType,
header.taskName, header.metricName, false,
improvementType).score }}
{{ getTaskScore(dataset, selectedType,
header.taskName, header.metricName, false,
improvementType).diffText }}
|