Batch Training - ICRL

Python

trajectories = await agent.train_batch(env_factory, goals)
results = await agent.run_batch(env_factory, eval_goals)

env_factory should return a fresh environment instance per goal.

TypeScript

const trajectories = await agent.trainBatch(envFactory, goals);
const results = await agent.runBatch(envFactory, evalGoals);

Suggested Workflow

Build a curriculum of goals.
Train in rounds.
Evaluate with run/runBatch on held-out goals.
Monitor DB stats and curation behavior.

Validation

Python DB includes deferred validation methods (validate_trajectory, validate_all) that can be run post-training for code-change persistence checks.