TrainingRunner

Training Runner - Training Process Controller

Responsible for:

• Intercepting Agent requests and determining if they should be used for training

• Replacing ChatClient to Trinity backend

• Collecting trace data

• Calculating rewards and calling feedback API

• Periodically calling commit API to trigger training

Usage example:

`// Method 1: Using sampling rate strategy
TrainingRunner runner = TrainingRunner.builder()
.trinityEndpoint("http://localhost:8080")
.modelName("/path/to/model")
.selectionStrategy(SamplingRateStrategy.of(0.1)) // 10% sampling rate
.rewardCalculator(agent -> 0.0) // Implement custom reward logic
.commitIntervalSeconds(300) // Commit every 5 minutes
.build();

// Method 2: Using explicit marking strategy
TrainingRunner runner = TrainingRunner.builder()
.trinityEndpoint("http://localhost:8080")
.modelName("/path/to/model")
.selectionStrategy(ExplicitMarkingStrategy.create()) // User manually marks
.rewardCalculator(agent -> {
// Custom reward calculation logic
return 1.0;
})
.build();

// Method 3: Using custom RewardCalculator class
TrainingRunner runner = TrainingRunner.builder()
.trinityEndpoint("http://localhost:8080")
.modelName("/path/to/model")
.selectionStrategy(SamplingRateStrategy.of(0.1))
.rewardCalculator(new CustomRewardCalculator()) // Implement RewardCalculator interface
.build();

// Start training runner
runner.start();

// Users use Agent normally (transparent)
agent.call(msg).block();

// Stop training
runner.stop();
`