Alpha Sigmoid Steepness

How the alpha sigmoid steepness subnet hyperparameter shapes how sharply liquid alpha responds to validator consensus alignment when smoothing validator-miner bonds in Yuma Consensus.

Alpha sigmoid steepness is a per-subnet hyperparameter, listed in the documentation as AlphaSigmoidSteepness, that shapes how alpha values are assigned based on validator voting alignment in Yuma Consensus. The Subnet Hyperparameters reference lists it as the control that determines how alpha values are assigned according to how a validator’s weights align with consensus.

What It Controls

The hyperparameter sets the steepness of the sigmoid curve that maps a validator’s consensus alignment to the alpha used when smoothing validator-miner bonds. It works together with the liquid alpha feature, which varies the bond-smoothing factor between an alpha_low and alpha_high bound rather than holding it constant (Consensus-based Weights).

A steeper curve makes the assigned alpha shift more sharply as a validator moves toward or away from consensus, while a gentler curve spreads that response over a wider alignment range. It does not set the alpha bounds themselves, which belong to the liquid alpha range.

Because the steepness only reshapes how alignment maps onto the smoothing factor, it cannot push that factor outside the low and high alpha bounds the liquid alpha range defines. At a very high steepness the sigmoid approaches a step response, so pairs just above the alignment midpoint snap toward the high bound while those just below snap toward the low bound, leaving little middle ground; a low steepness keeps the transition gradual, so small alignment differences produce only small changes in the smoothing factor (Consensus-based Weights, Subnet Hyperparameters).

Documented Type, Default, and Setter

The Subnet Hyperparameters reference lists AlphaSigmoidSteepness as an i16 with a default of 1000. Because each subnet carries its own hyperparameters, the steepness is defined per subnet and applies only within that subnet’s bond-smoothing path.

That same default of 1000 is exposed as a chain startup constant, listed in the constants reference as api.consts.subtensorModule.initialAlphaSigmoidSteepness with the raw value 1000. That constant is the steepness the chain seeds a new subnet with at launch, while the live AlphaSigmoidSteepness storage value is the per-subnet figure used afterward, so a subnet’s current steepness can differ from this initial constant once it has been changed through the setter described below (Subtensor Constants: initialAlphaSigmoidSteepness, Subnet Hyperparameters).

It is changed through the sudo_set_alpha_sigmoid_steepness extrinsic, exposed in tooling as btcli sudo set --param alpha_sigmoid_steepness. In the current btcli mapping this parameter is marked as not owner-settable and root-sudo-only, so changing it requires a root sudo call rather than an ordinary subnet-owner hyperparameter update (Subnet Hyperparameters).

The reference describes the parameter itself as determining how the consensus mechanism assigns an alpha value for a given miner-validator pair based on voting alignment, which is why it sits on the bond-smoothing path rather than on registration or emission directly (Subnet Hyperparameters).

The setter writes only this steepness; it does not move the alpha_low and alpha_high bounds, which the reference documents as a separate AlphaValues setting. Steepness therefore reshapes how sharply the assigned alpha travels between those endpoints, but the endpoints themselves remain fixed until AlphaValues is changed (Subnet Hyperparameters).

The reference also fixes the direction of the control: it describes lower steepness values as producing moderate alpha values, while higher steepness values push the assigned alpha closer to the defined alpha_low or alpha_high bounds, depending on which side of consensus a miner-validator pair falls on (Subnet Hyperparameters).

The value in force is per-subnet chain state and can differ from the documented default, so a subnet’s real steepness should be read for that subnet rather than assumed from the reference value. For the documented example netuid 1 on Finney mainnet, the live value can be read from that subnet’s hyperparameter table with btcli subnet hyperparameters --netuid 1 --network finney, which keeps the steepness claim reproducible against a parseable netuid (Subnet Hyperparameters: View hyperparameters).

Distinction from Liquid Alpha

Liquid alpha is the feature that varies the bond-smoothing factor by consensus alignment, while alpha sigmoid steepness sets how sharp that variation is across the alignment range (Subnet Hyperparameters, Consensus-based Weights).

  • Liquid alpha — enables a variable, consensus-aligned bond-smoothing factor.
  • Alpha sigmoid steepness — sets how steeply that factor responds to alignment.

Distinction from Bonds Moving Average

Alpha sigmoid steepness and bonds moving average both touch the bond path but answer different questions. Steepness shapes the response curve that assigns alpha from consensus alignment, while bonds moving average sets the moving-average behavior applied to validator-miner bonds over time (Subnet Hyperparameters, Yuma Consensus).

  • Alpha sigmoid steepness — the sharpness of the alpha-assignment curve.
  • Bonds moving average — the smoothing applied to bonds across epochs.

Distinction from Alpha Low and Alpha High

Alpha sigmoid steepness works within the range that alpha low and alpha high define. Alpha low and alpha high fix the lower and upper endpoints of the bond-smoothing coefficient, while alpha sigmoid steepness sets how sharply the coefficient moves between those endpoints (Subnet Hyperparameters, Consensus-based Weights).

The bounds set where the coefficient can land; steepness sets the sharpness of its movement within that range.

  • Alpha low / alpha high — the lower and upper endpoints of the coefficient range.
  • Alpha sigmoid steepness — how sharply the coefficient moves within that range.

Distinction from EMA

The alpha that the steepness curve assigns is the EMA smoothing coefficient used when updating validator-miner bonds. EMA is the technique that blends each epoch’s new bond value with the prior one, while alpha sigmoid steepness shapes how the coefficient feeding that EMA is set from a validator’s consensus alignment under the liquid alpha feature (EMA, Consensus-based Weights, Glossary: Validator-Miner Bonds).

  • EMA — the moving-average technique that blends new and prior bond values.
  • Alpha sigmoid steepness — shapes the coefficient that EMA uses under liquid alpha.

Chain Reads for netuid 1

Readers can verify live hyperparameter values for the documented example netuid with btcli subnet hyperparameters --netuid 1 --network finney (Subnet Hyperparameters: View hyperparameters).

That read path keeps live hyperparameter claims tied to a parseable netuid.

Per-Subnet Live Value Boundary

Alpha sigmoid steepness is per-subnet chain state. The steepness value configured on one netuid shapes how sharply liquid alpha responds inside that subnet’s bond-smoothing path, but the live value can differ from reference defaults and from steepness on another netuid (Subnet Hyperparameters, Consensus-based Weights).

This article’s infobox uses netuid 1 as an example label when reading bond-path settings on one subnet. That example helps readers verify one table; it is not proof that every subnet exposes the same live steepness (Bittensor Networks).

Alpha sigmoid steepness governs bond-path response shape when liquid alpha is enabled; Yuma Consensus still settles validator weights into emission shares each epoch (Yuma Consensus).

  • Alpha sigmoid steepness — liquid-alpha response steepness inside bond smoothing on a subnet.
  • Per-subnet boundary — live value is chain state on one netuid.

Distinction from Yuma Consensus

Alpha sigmoid steepness is a per-subnet hyperparameter that shapes how sharply the liquid-alpha coefficient responds to a validator’s consensus alignment inside Yuma Consensus bonding. Yuma Consensus is the separate tempo process that aggregates eligible validator weights, applies clipping, updates bonds, and maps the result into emission shares each epoch (Subnet Hyperparameters, Consensus-based Weights, Yuma Consensus, Emission).

  • Alpha sigmoid steepness — sharpness of the liquid-alpha response curve inside bonding.

  • Yuma Consensus — aggregation that turns filtered weights and bond state into emission shares.

  • Liquid alpha — feature that varies the bond-smoothing coefficient by consensus alignment.

  • Bond-path control — steepness tunes how alignment maps into bond adaptation, not miner task ranking (Consensus-based Weights, Yuma Consensus).

  • Requires liquid alpha — steepness matters only when liquid alpha is enabled on the subnet (Glossary: Liquid Alpha).

Reader Boundary

This page defines alpha sigmoid steepness at a high level. It does not report the live steepness value on any particular subnet or whether liquid alpha is enabled there. Those are per-subnet hyperparameter settings that should be checked for the relevant netuid (Subnet Hyperparameters).

Further Reading

Topics ConsensusSubnets
Last updated (UTC) 7 min read