EvolutionaryScale · Schema

EvolutionaryScale GenerationConfig

Schema for the `GenerationConfig` object that controls iterative masked sampling for the EvolutionaryScale Forge ESM3 generate endpoint.

AIArtificial IntelligenceBiologyBioinformaticsComputational BiologyDrug DiscoveryESMESM3ESM CambrianFoundation ModelsGenerative BiologyLife SciencesMachine LearningProtein DesignProtein FoldingProtein Language ModelsProteinsRepresentation LearningStructure Prediction

Properties

Name	Type	Description
track	string	Which ESM3 track to generate. Sequence is the most common; structure and function tracks require the multimodal model checkpoints.
num_steps	integer	Number of iterative masked-sampling steps. Lower values are faster; higher values produce more refined outputs.
temperature	number	Sampling temperature. 0.0 yields greedy decoding; 1.0 matches the trained distribution; higher values increase diversity.
top_p	number	Nucleus sampling cutoff. Samples from the smallest set of tokens whose cumulative probability is at least `top_p`.
schedule	string	Decoding schedule controlling how many positions are unmasked per step.
invalid_ids	array	Token IDs the sampler must never emit. Useful for constraining amino acid alphabets or forbidding rare structure tokens.
condition_on_coordinates_only	boolean	If true, condition only on coordinates (ignore sequence tokens) when generating other tracks.

View JSON Schema on GitHub

JSON Schema

{
  "$schema": "https://json-schema.org/draft/2020-12/schema",
  "$id": "https://api-evangelist.com/schemas/evolutionaryscale/evolutionaryscale-generation-config-schema.json",
  "title": "EvolutionaryScale GenerationConfig",
  "description": "Schema for the `GenerationConfig` object that controls iterative masked sampling for the EvolutionaryScale Forge ESM3 generate endpoint.",
  "type": "object",
  "required": ["track"],
  "properties": {
    "track": {
      "type": "string",
      "description": "Which ESM3 track to generate. Sequence is the most common; structure and function tracks require the multimodal model checkpoints.",
      "enum": ["sequence", "structure", "secondary_structure", "sasa", "function"]
    },
    "num_steps": {
      "type": "integer",
      "description": "Number of iterative masked-sampling steps. Lower values are faster; higher values produce more refined outputs.",
      "default": 8,
      "minimum": 1,
      "maximum": 256
    },
    "temperature": {
      "type": "number",
      "description": "Sampling temperature. 0.0 yields greedy decoding; 1.0 matches the trained distribution; higher values increase diversity.",
      "default": 1.0,
      "minimum": 0.0
    },
    "top_p": {
      "type": "number",
      "description": "Nucleus sampling cutoff. Samples from the smallest set of tokens whose cumulative probability is at least `top_p`.",
      "minimum": 0.0,
      "maximum": 1.0
    },
    "schedule": {
      "type": "string",
      "description": "Decoding schedule controlling how many positions are unmasked per step.",
      "enum": ["cosine", "linear"]
    },
    "invalid_ids": {
      "type": "array",
      "description": "Token IDs the sampler must never emit. Useful for constraining amino acid alphabets or forbidding rare structure tokens.",
      "items": { "type": "integer", "minimum": 0 }
    },
    "condition_on_coordinates_only": {
      "type": "boolean",
      "description": "If true, condition only on coordinates (ignore sequence tokens) when generating other tracks.",
      "default": false
    }
  },
  "additionalProperties": false
}