wehub-resource-sync
94 lines
3.2 KiB
Protocol Buffer
94 lines
3.2 KiB
Protocol Buffer
syntax = "proto2";
|
|
|
|
package mlflow;
|
|
|
|
import "jobs.proto";
|
|
import "scalapb/scalapb.proto";
|
|
|
|
option java_package = "org.mlflow.api.proto";
|
|
option py_generic_services = true;
|
|
option (scalapb.options) = {flat_package: true};
|
|
|
|
// Type of optimizer algorithm to use.
|
|
enum OptimizerType {
|
|
OPTIMIZER_TYPE_UNSPECIFIED = 0;
|
|
|
|
// GEPA (Genetic Pareto) optimizer (https://github.com/gepa-ai/gepa)
|
|
OPTIMIZER_TYPE_GEPA = 1;
|
|
|
|
// MetaPrompt optimizer - uses metaprompting with LLMs to improve prompts in a single pass.
|
|
OPTIMIZER_TYPE_METAPROMPT = 2;
|
|
}
|
|
|
|
// Tag for a prompt optimization job.
|
|
message PromptOptimizationJobTag {
|
|
optional string key = 1;
|
|
|
|
optional string value = 2;
|
|
}
|
|
|
|
// Configuration for a prompt optimization job.
|
|
// Stored as run parameters in the underlying MLflow run.
|
|
message PromptOptimizationJobConfig {
|
|
// The optimizer type to use.
|
|
optional OptimizerType optimizer_type = 1;
|
|
|
|
// ID of the EvaluationDataset containing training data.
|
|
optional string dataset_id = 2;
|
|
|
|
// List of scorer names. Can be built-in scorer class names
|
|
// (e.g., "Correctness", "Safety") or registered scorer names.
|
|
repeated string scorers = 3;
|
|
|
|
// JSON-serialized optimizer-specific configuration.
|
|
// Different optimizers accept different parameters:
|
|
// - GEPA: {"reflection_model": "openai:/gpt-5", "max_metric_calls": 300}
|
|
// - MetaPrompt: {"reflection_model": "openai:/gpt-5", "guidelines": "...", "lm_kwargs": {...}}
|
|
optional string optimizer_config_json = 4;
|
|
}
|
|
|
|
// Represents a prompt optimization job entity.
|
|
message PromptOptimizationJob {
|
|
// Unique identifier for the optimization job.
|
|
// Used to poll job execution status (pending/running/completed/failed).
|
|
optional string job_id = 1;
|
|
|
|
// MLflow run ID where optimization metrics and results are stored.
|
|
// Use this to view results in MLflow UI. Only available after job starts running.
|
|
optional string run_id = 2;
|
|
|
|
// Current state of the job (status + error message + metadata).
|
|
optional JobState state = 3;
|
|
|
|
// ID of the MLflow experiment where this optimization job is tracked.
|
|
optional string experiment_id = 4;
|
|
|
|
// URI of the source prompt that optimization started from (e.g., "prompts:/my-prompt/1").
|
|
optional string source_prompt_uri = 5;
|
|
|
|
// URI of the optimized prompt (e.g., "prompts:/my-prompt/2").
|
|
// Only set if optimization completed successfully.
|
|
optional string optimized_prompt_uri = 6;
|
|
|
|
// Configuration for the optimization job.
|
|
optional PromptOptimizationJobConfig config = 7;
|
|
|
|
// Timestamp when the job was created (milliseconds since epoch).
|
|
optional int64 creation_timestamp_ms = 8;
|
|
|
|
// Timestamp when the job completed (milliseconds since epoch).
|
|
// Only set if status is COMPLETED, FAILED, or CANCELED.
|
|
optional int64 completion_timestamp_ms = 9;
|
|
|
|
// Tags associated with this job.
|
|
repeated PromptOptimizationJobTag tags = 10;
|
|
|
|
// Initial evaluation scores before optimization, keyed by scorer name.
|
|
// Example: {"Correctness": 0.65, "Safety": 0.80}
|
|
map<string, double> initial_eval_scores = 11;
|
|
|
|
// Final evaluation scores after optimization, keyed by scorer name.
|
|
// Example: {"Correctness": 0.89, "Safety": 0.95}
|
|
map<string, double> final_eval_scores = 12;
|
|
}
|