Configuration Reference

All config lives in YAML files. configs/base.yaml is the default; experiment configs override it via base: configs/base.yaml.

Top-Level

Key	Type	Default	Description
`mode`	str	`"submit"`	`"submit"` for Kaggle inference, `"train"` for local training

Key	Type	Default	Description
`data_dir`	str	`/kaggle/input/competitions/birdclef-2026`	Competition data root
`model_dir`	str	`/kaggle/input/models/.../perch_v2_cpu/1`	Perch v2 model directory
`cache_dir`	str	`cache/`	Writable embedding cache directory
`cache_input_dir`	str	`/kaggle/input/perch-meta`	Read-only pre-computed cache dataset

Key	Type	Default	Description
`batch_files`	int	`32`	Audio files per Perch embedding batch
`proxy_reduce`	str	`"max"`	Reduction over proxy scores (`"max"` or `"mean"`)
`require_full_cache_in_submit`	bool	`true`	Fail fast if cache missing in submit mode
`verbose`	bool	`false`	Extra logging
`run_oof_baseline`	bool	`false`	Compute OOF baseline scores
`run_probe_check`	bool	`false`	Sanity-check probe predictions
`dryrun_n_files`	int	`20`	Files to process in dryrun/debug mode

These are fixed and should not be tuned.

Key	Type	Default	Description
`d_model`	int	`128`	Hidden dimension
`d_state`	int	`16`	SSM state size
`n_ssm_layers`	int	`2`	Number of SSM layers
`dropout`	float	`0.15`	Dropout rate
`n_prototypes`	int	`1`	Prototypes per class
`n_sites`	int	`20`	Site-embedding cardinality
`meta_dim`	int	`16`	Metadata embedding dimension
`use_cross_attn`	bool	`true`	Enable cross-attention layer
`cross_attn_heads`	int	`4`	Number of cross-attention heads

Parameters marked with a star are included in the neuropt search space.

Key	Type	Default	neuropt	Description
`n_epochs`	int	`35`	--	Training epochs
`lr`	float	`5.5e-4`	log_uniform [5e-4, 3e-3]	Learning rate
`weight_decay`	float	`1e-3`	log_uniform [8e-4, 5e-3]	AdamW weight decay
`val_ratio`	float	`0.15`	--	Validation holdout fraction
`patience`	int	`15`	--	Early-stopping patience (epochs)
`pos_weight_cap`	float	`41.0`	uniform [20, 45]	Cap on positive class weight
`distill_weight`	float	`0.23`	uniform [0.05, 0.25]	Distillation loss weight
`proto_margin`	float	`0.1`	--	Prototype margin loss
`label_smoothing`	float	`0.019`	uniform [0.01, 0.05]	Label smoothing factor
`oof_n_splits`	int	`3`	--	OOF cross-validation folds
`mixup_alpha`	float	`0.25`	uniform [0.1, 0.45]	Mixup interpolation alpha
`focal_gamma`	float	`1.15`	uniform [1.0, 3.0]	Focal loss gamma
`swa_start_frac`	float	`0.68`	uniform [0.6, 0.85]	Fraction of training before SWA
`swa_lr`	float	`5e-4`	--	SWA learning rate

Key	Type	Default	Description
`lambda_event`	float	`0.4`	Weight for event-level scores
`lambda_texture`	float	`1.0`	Weight for texture-level scores
`lambda_proxy_texture`	float	`0.8`	Weight for proxy-texture blend
`smooth_texture`	float	`0.35`	Temporal smoothing for texture stream
`smooth_event`	float	`0.15`	Temporal smoothing for event stream

Key	Type	Default	Description
`d_model`	int	`64`	Hidden dimension
`d_state`	int	`8`	SSM state size
`n_ssm_layers`	int	`1`	Number of SSM layers
`dropout`	float	`0.1`	Dropout rate
`correction_weight`	float	`0.3`	Blend weight for residual correction
`n_epochs`	int	`20`	Training epochs
`lr`	float	`1e-3`	Learning rate
`patience`	int	`8`	Early-stopping patience

Key	Type	Default	Description
`aves`	float	`1.10`	Temperature for bird (Aves) logits
`texture`	float	`0.95`	Temperature for texture taxa (Amphibia, Insecta)

Key	Type	Default	Description
`file_level_top_k`	int	`2`	Keep top-k species per file
`tta_shifts`	list[int]	`[0]`	Test-time augmentation shifts (seconds)

Key	Type	Default	Description
`backend`	str	`"mlp"`	`"mlp"` or `"logistic"`
`pca_dim`	int	`64`	PCA dimension for probe input
`min_pos`	int	`8`	Minimum positive examples to train a probe
`C`	float	`0.5`	Logistic regularisation strength
`alpha`	float	`0.4`	Probe blend weight

Key	Type	Default	Description
`hidden_layer_sizes`	list[int]	`[128]`	MLP hidden layers
`activation`	str	`"relu"`	Activation function
`max_iter`	int	`100`	Max training iterations
`early_stopping`	bool	`true`	Enable early stopping
`validation_fraction`	float	`0.15`	Validation split for early stopping
`n_iter_no_change`	int	`10`	Early stopping patience
`learning_rate_init`	float	`0.001`	Initial learning rate
`l2_alpha`	float	`0.01`	L2 regularisation

Key	Type	Default	Description
`budget_seconds`	float	`5400.0`	Total wall-clock budget (90 min)
`residual_ssm_min_remaining`	float	`240.0`	Skip residual SSM if less time remains

Key	Type	Default	Description
`search_space`	dict	(see below)	Parameter search ranges
`batch_size`	int	`3`	Parallel evaluations per neuropt round
`max_evals`	int	`60`	Maximum total evaluations
`backend`	str	`"claude"`	neuropt backend