[VictorMorand]

aimed for configurations: current example:

  pre_experiment: Mocking (not importing) the following modules in experiment manager: ['torch', 'torchmetrics', 'torchdata', 'lightning', 'pytorch_lightning', 'lightning_fabric', 'sentence_transformers', 'transformers', 'safetensors.torch', 'pylate', 'xpmir.learning.losses', 'xpmir.neural._sparton', 'xpm_torch.datasets', 'MICE.modeling.mice_v2.modernbert_hf']
2026-10-08T22:20:57 WARN  experimaestro.utils.config:
================================================================================
⚠️  WARNING: Unrecognized configuration keys for BaselinesConfig:
   Unknown keys: ['ettin_subset_exclude', 'learner', 'validation']
   Valid keys are: ['add_timestamp', 'compare_with_baseline', 'description', 'dirty_git', 'doc_prompt', 'eval_sparsity', 'evaluation', 'file', 'from_experiment', 'from_run_id', 'gpu', 'grid_search', 'hard_prune_threshold', 'hard_pruning', 'id', 'imports', 'include_recall_100', 'indexation', 'max_doc_length', 'max_length', 'max_query_length', 'module', 'paper', 'parent', 'plaid', 'pre_experiment', 'precision', 'precompute_first_stage', 'pref_attn_implementation', 'preprocessing', 'pythonpath', 'query_prompt', 'retrieval', 'retriever', 'save_runs', 'scorer_id', 'scorer_type', 'seed', 'seismic', 'subtitle', 'title', 'use_best_device', 'use_plaid']
================================================================================

[bpiwowar]

Wouldn't it be better to fail in those cases (it was the case for simple configurations => mismatching keys aborts experiments).

[VictorMorand]

ok I agree - fixed this to use strict keys matching in configurations, also fixed tests now