Search parameters reference

Every parameter accepted under the search block, for reward guided (beam) search. This page is a lookup table; see Reward guided search for how the pieces fit together, when to use use_clean_prediction, and the novelty-against-passers walkthrough. See the CLI options reference for the handful of search.* keys you can also override directly on the command line.

Contents

• Top-level search block

• Per-checkpoint parameters

• Reward terms

• continuation_criteria and save_criteria thresholds

• Novelty metric parameters

Top-level search block

Parameter	Type	Default	Description
beam_width	int	4	Candidates retained per input after every scored checkpoint.
n_branch	int	4	Default number of children per surviving beam, for any checkpoint interval that doesn’t set its own n_branch.
save_rollouts	bool	false	Also return completed scoring rollouts that lost selection, alongside the final beam. Required to set save_criteria.
save_criteria	mapping of metric key → threshold	{}	Raw-metric thresholds (see thresholds) an extra rollout must pass to be saved. Every checkpoint that can produce a saved rollout must list the providers these keys need. Can reference search.reward.
checkpoints	mapping of step → checkpoint config	Required, nonempty	Positive integer integration steps, see below. Quote YAML keys to support dotted CLI overrides such as ++search.checkpoints.20.use_clean_prediction=true. Step zero is implicit and cannot be listed.

Per-checkpoint parameters

Each entry in search.checkpoints is keyed by a positive integer step and accepts exactly these fields. Nothing carries forward from one checkpoint to the next — every field is self-contained.

Parameter	Type	Default	Description
metrics	list of metric names	[]	Named providers from metrics.inline_metrics, metrics.refolding_oracles, or metrics.refolding_metrics to run at this checkpoint. Runs regardless of dataset selection tags; selecting a provider computes all of its outputs. A checkpoint at the schedule end must list providers that score completed structures (no oracle/refolding/pass-criteria providers if use_clean_prediction: true).
rewards	mapping of metric.output_key → reward term	{}	Chooses which of the selected metrics’ outputs contribute to ranking, and how. An empty mapping gives every candidate zero reward, preserving deterministic input order at selection.
n_branch	positive int	the global search.n_branch	Overrides the branch count for this checkpoint’s incoming interval only.
use_clean_prediction	bool	false	Score the clean (decoded) prediction from the preceding integration step instead of completing a separate scoring rollout. Not valid on the schedule-end checkpoint. Rejects oracle, refolding, and pass-criteria providers in metrics. Predictions are used only for ranking — never saved as completed designs or sent to refolding.
continuation_criteria	mapping of metric key → threshold	{}	Raw-metric thresholds that decide whether a checkpoint before the schedule end stops an input. The whole selected beam continues if any selected candidate passes every threshold; it stops only if all selected candidates fail. Every metric a criterion needs must also appear in this checkpoint’s metrics. Not valid on the schedule-end checkpoint.

Reward terms

Each entry in a checkpoint’s rewards mapping is keyed metric_name.output_key and accepts:

Field	Type	Default	Description
weight	finite number	1.0	Multiplier applied after clipping and normalizing. Negative weights minimize the term. A zero weight contributes nothing to ranking or selection.
clip	mapping with min and/or max	no clipping	Clips the raw metric value before normalizing. At least one of min/max must be set; min must not exceed max.
normalize_by	positive finite number	1.0	Divides the (clipped) value before weighting.

Terms are summed per candidate into search.reward. Clipping and normalization affect the reward only — the underlying metric values stay raw wherever else they’re reported.

continuation_criteria and save_criteria thresholds

Both fields use the same threshold schema, keyed by a metric output (or search.reward) or by an arbitrary label when reducing several keys:

Field	Type	Description
min	number	Candidate value must be at least this (inclusive).
max	number	Candidate value must be at most this (inclusive).
min_of	list of metric keys	Reduce several keys to their minimum before thresholding with min/max. Mutually exclusive with max_of/mean_of.
max_of	list of metric keys	Reduce several keys to their maximum before thresholding.
mean_of	list of metric keys	Reduce several keys to their mean before thresholding.

All listed thresholds must pass. A metric missing or non-finite for a candidate fails that candidate; a metric absent from every candidate in the pool is a configuration error.

Novelty metric parameters

FoldseekNoveltyMetric (rfproteina.metrics.foldseek_novelty.FoldseekNoveltyMetric) is an ordinary metric you select like any other at a checkpoint, then gate with continuation_criteria on its is_novel output. See Novelty against passing designs.

Parameter	Type	Default	Description
archive_dir	path	Required	Append-only directory of archived passer CIFs, populated by a PassCriterionMetric with its own archive_dir set. Use one archive per target and pass definition.
tm_threshold	float in (0, 1]	0.5	Minimum query- and target-normalized TM score for a candidate to count as a match against an archived passer.
min_coverage	float in (0, 1]	0.8	Minimum aligned fraction of both query and target for a match.
threads	positive int	1	CPU threads per Foldseek invocation.
timeout	positive float (seconds)	300.0	Maximum time for each Foldseek invocation; a timeout raises rather than reporting novelty.

Output: is_novel is 0 on any match, 1 if there is no match or the archive is empty. Requires the FOLDSEEK_EXEC environment variable pointing at the Foldseek binary. Each metric instance keeps a local database copy and rebuilds it only when the archive’s published files change.