Configuration

An experiment is a YAML file that names a profile and a matrix:

# configs/experiments/cpu_main.yaml
name: cpu_main
profile: cpu
matrix:
  preset: [split_mnist_class, split_mnist_task, permuted_mnist, split_cifar10_class, split_cifar100_class, text5_class]
  method.name: [naive, joint, ewc, lwf, er, derpp, agem, lora_per_task,
                'consolidate:ta', 'consolidate:ties', 'consolidate:distill', 'consolidate:continual_distill']
  seed: [0, 1, 2]
workers: 2
threads_per_worker: 2

A profile (configs/profiles/{smoke,cpu,gpu}.yaml) holds defaults and named presets (benchmark + model + training scale).

Merge order

Later entries win:

  1. dataclass defaults (contilab/core/config.py)
  2. profile.defaults
  3. the preset
  4. experiment.overrides
  5. matrix axes (dotted paths such as method.name, consolidate.transfer)
  6. sweep values
  7. command-line key=value

Unknown keys are errors.

Command-line overrides 

contilab run configs/experiments/cpu_main.yaml train.epochs=2 optim.lr=0.01
contilab run configs/experiments/cpu_main.yaml "matrix.preset=[split_mnist_class]" "matrix.seed=[0]"

matrix.<axis>=[...] replaces a matrix axis.

Sweeps and variants 

sweep:
  method.params.lam: [1, 10, 100]

Any extra matrix axis or sweep key becomes a variant column in results.csv and its own row in report tables.

Methods usually have different hyperparameters, so method_sweep gives each method its own grid; methods without an entry run once:

method_sweep:
  ewc: {method.params.lam: [1, 3, 10, 30, 100, 300]}
  derpp: {method.params.alpha: [0.1, 0.3, 1.0], method.params.beta: [0.5, 1.0]}
  'consolidate:ties': {consolidate.ties_k: [0.1, 0.2, 0.5]}

Validation-based selection

Every run records the final model’s accuracy on each task’s validation split (10 % of its training data, never used for training) as val_acc. For each (preset, method) with several variants, results.csv marks the variant with the best mean val_acc over seeds in a selected column. Diverged seeds score 0, so unstable settings lose. Only hyperparameters (sweep, method_sweep) are selected. Extra matrix axes such as consolidate.transfer are comparison axes, so selection runs separately for each of their values. Reports tag the selected variant with ✓. The cross-benchmark ranking uses only selected variants, labelled <method> (tuned), so a method keeps one row even when its best setting differs between benchmarks. Test-set metrics never influence the choice.

Main keys

Section Keys (default)
benchmark name, scenario (class), n_train, n_test, n_tasks, max_len (128), tokenizer, task_order_seed
model name (mlp | resnet18r | hf), hidden (400), depth (2), nf (20), hf_name, lora.{r: 8, alpha: 16, targets, lr: 1e-3}
optim name (sgd | adamw), lr (0.03), weight_decay, momentum, clip
train epochs (1), batch_size (32), eval_batch_size (256)
engine device (auto), compile (auto), precision (auto), max_E (8)
method name, params (strategy-specific, merged over per-modality defaults)
consolidate lam_grid (0.1…1.0), ties_k (0.2), init (ties), routing (routed), transfer (buffer), M (500), temperature (2), epochs (10), experts_lora (false), init_seed (0)
metrics tier (core | extended | full), rho (50), stream_subset (200), beta (10), probe_size (256), recovery_n (50), recovery_steps (20), window_frac (0.1), ece_bins (15), dormant_tau (0.025), srank_delta (0.01), ppl_n (200), barrier_points (11)