| """Real source-data partition checks; no model/editor inference.""" |
| from collections import Counter |
| from copy import deepcopy |
| import json |
| from pathlib import Path |
| import pytest |
| from iwibench.qualified_catalog import catalog,qualified_episode |
| from iwibench.task_partitions import partitions,partition_indices,meaningful_digest |
| from iwibench.taskset import IWIBenchConfig,IWIBenchTaskset |
| from iwibench.variants import task_from_row |
|
|
| def test_full_coverage_and_stratified_disjoint_meaningful_groups(): |
| part=partitions();groups={s:[x for x in part['entries'] if x['dataset_split']==s] for s in ('train','test')} |
| assert Counter(x['family'] for x in groups['train'])==dict.fromkeys(catalog()['counts'],85) |
| assert Counter(x['family'] for x in groups['test'])==dict.fromkeys(catalog()['counts'],15) |
| assert len({x['id'] for x in part['entries']})==1000 |
| for key in ('id','parameter_sha256','episode_sha256','meaningful_sha256'): |
| assert not {x[key] for x in groups['train']}&{x[key] for x in groups['test']} |
| for entry in part['entries']: |
| assert meaningful_digest(qualified_episode(entry['family'],entry['variant_index']))==entry['meaningful_sha256'] |
|
|
| def test_no_test_task_equivalent_to_published_history(): |
| root=Path(__file__).resolve().parents[1];hist=json.loads((root/'data/historical_task_overlap.json').read_text());assert len(hist['rows'])==325 |
| old={x['meaningful_sha256'] for x in hist['rows']};test={x['meaningful_sha256'] for x in partitions()['entries'] if x['dataset_split']=='test'};assert not old&test |
| assert sum(bool(x['historical_equivalent_archive_episode_ids']) for x in partitions()['entries'])==25 |
|
|
| def test_native_loader_filter_preserves_original_ids_and_seeds(): |
| sets={} |
| for split,n in [('train',850),('test',150),('all',1000)]: |
| ts=list(IWIBenchTaskset(IWIBenchConfig(model_id='google/gemma-3-4b-it',variant_limit=100,variant_catalog='baseline-qualified-v1',dataset_partition=split))) |
| assert len(ts)==n |
| sets[split]={x.key for x in ts} |
| for task in ts:assert task.data.seed==qualified_episode(task.data.family,task.data.variant_index)['seed'] and task.data.split=='development' |
| assert not sets['train']&sets['test'] and sets['train']|sets['test']==sets['all'] |
| with pytest.raises(ValueError):IWIBenchConfig(dataset_partition='test') |
|
|
| def test_semantic_grouping_ignores_annotation_ids_and_seed_padding(): |
| ep=qualified_episode('factual_update',0);changed=deepcopy(ep);changed['id']='different';changed['seed']+=1;changed['parameters']['fact_id']='different_annotation_id';changed['parameters']['stale_candidates']=['different former holder'] |
| assert meaningful_digest(ep)==meaningful_digest(changed) |
| changed['parameters']['entity']='different institution';assert meaningful_digest(ep)!=meaningful_digest(changed) |
| ep=qualified_episode('resistance_to_false_suggestions',0);changed=deepcopy(ep);changed['parameters']['problem_seed']+=100000;assert meaningful_digest(ep)==meaningful_digest(changed) |
| changed['targets'][0]['answer']+=1;assert meaningful_digest(ep)!=meaningful_digest(changed) |
|
|
| def test_rollout_schema_is_not_silently_accepted_as_task(): |
| with pytest.raises(ValueError,match='not editing tasks'):task_from_row({'archive_episode_id':'real-rollout-index-row'}) |
|
|
| def test_partition_index_validation_and_defensive_lists(): |
| with pytest.raises(ValueError):partition_indices('alias_acquisition','validation') |
| indices=partition_indices('alias_acquisition','test');indices.clear();assert len(partition_indices('alias_acquisition','test'))==15 |
|
|
