rleaas 1.1.0

Release SDK — RL Environments as a Service by Centific

rleaas — Release SDK

Python SDK for the Release (RLEaaS) platform by Centific — RL Environments as a Service.

Installation

pip install rleaas

API Key

Setup	API key needed?	How to get it
Local dev server (http://localhost:8000)	No	Run the AgentWork-Simulator locally — no auth required
Hosted / production	Yes	Log in to the Release dashboard → Settings → API Keys → create a key

Set the key as an environment variable so you never hardcode it:

export RLEAAS_API_KEY="rleaas_sk_your_key_here"

The SDK reads it automatically:

import rleaas

client = rleaas.Client()                          # reads RLEAAS_API_KEY from env
# or
client = rleaas.Client(api_key="rleaas_sk_...")   # pass explicitly

Local development (no key required)

import rleaas

client = rleaas.Client(base_url="http://localhost:8000")
print(client.ping())
# {'message': 'RL Environment & Agent API', 'version': '1.0.0', ...}

Sub-clients

Attribute	Purpose
client.Environment	Create and manage simulation environments
client.Tools	Register and configure agent tools
client.Agent	Register and export trained agents
client.Verifier	Define scoring verifiers (rule-based, LLM judge, composite)
client.Scenario	Create and browse training scenarios
client.ScenarioSuite	Organize scenarios into training/evaluation suites
client.TrainingJob	Launch and monitor GRPO/PPO/DQN/A2C training runs
client.Evaluation	Run evaluations and retrieve rollouts
client.Metrics	Query KPIs and training metrics
client.AuditLog	Access audit logs and governance configuration

Example

import rleaas

client = rleaas.Client()   # reads RLEAAS_API_KEY from environment

# Create environment
env = client.Environment.create(name="FinSim-Prod-v1", vertical="FinSim")
env.wait_until_ready()

# Create verifier
rule_v = client.Verifier.create(
    name="AML Compliance Check",
    verifier_type="rule_based",
    environment="FinSim-Prod-v1",
    config={
        "conditions": ["'run_aml_check' in trajectory.tool_calls"],
        "condition_logic": "AND",
        "reward_on_pass": 1.0,
        "reward_on_fail": 0.0,
    },
)

# Train
job = client.TrainingJob.run(
    environment_name="FinSim-Prod-v1",
    algorithm="GRPO",
    config={"episodes": 10000, "max_steps_per_episode": 20},
    verifier_ids=[rule_v.id],
)
job.wait_until_complete()
best = job.get_best_checkpoint()

# Evaluate
eval_job = client.Evaluation.run(
    agent_checkpoint_id=best["id"],
    scenario_suite_id="suite_eval_01",
    verifier_ids=[rule_v.id],
)
report = eval_job.wait_until_complete()
print(report["overall_score"])

Training + Simulation configuration

Use this simple flow:

1. Put training settings in your config.json under training as an array of entries.
2. Run python -m rleaas.training_cli commands to start and manage jobs.

Quick helper option:

python -m rleaas.training_cli start
# status / rerun / operations:
python -m rleaas.training_cli status --job-id <job_id>
Eg: 
python -m rleaas.training_cli status --job-id train_8f3a2b1c
python -m rleaas.training_cli list
python -m rleaas.training_cli list --ids-only
python -m rleaas.training_cli wait --job-id <job_id>
python -m rleaas.training_cli metrics --job-id <job_id>
python -m rleaas.training_cli checkpoints --job-id <job_id>
python -m rleaas.training_cli rollouts --job-id <job_id>
python -m rleaas.training_cli cancel --job-id <job_id>

# if config has multiple training entries:
python -m rleaas.training_cli start --training 3
python -m rleaas.training_cli start --training 3,4,5
python -m rleaas.training_cli start --all

config.json training schema

Place training config under top-level training key in config.json using this format only:

• "training": [ { ... }, { ... } ]

• each entry must include a training identifier (for example: 1, 2, 3)

• training (required, integer/string): run identifier (1, 2, 3, etc.).

• environment_name (required, string): target environment name to train.

• name (optional, string): run name shown in training history/UI.

• description (optional, string): objective/notes for the run.

• agent_id (optional, string): agent/model id to use for training.

• scenario_id (optional, string): scenario/suite id for the run.

• verifier_ids (optional, array of strings): verifier ids to evaluate rewards/scoring.

• algorithm (optional, string, default PPO): one of GRPO, PPO, SAC, DQN, A2C, A3C, TD3, DDPG, SLM.

• max_steps (optional, integer, default 200 in README flow): max steps per episode.

• episodes (optional, integer, default 100 in README flow): number of episodes.

• reward_fn (optional): reward function reference in one of these formats:

  • string: "rewards/finsim_reward.py" or inline expression/function string
  • object path form: { "path": "rewards/finsim_reward.py" }
  • object inline form: { "inline": "reward = ..." }

• simulation (optional, object):

  • speed (optional, positive number)
  • seed (optional, non-negative integer)
  • episode_settings (optional, object):
    • num_episodes (optional, positive integer)
    • max_steps (optional, positive integer)

• config (optional, object): additional backend config, for example learning_rate, batch_size, checkpoint_interval.

Example:

{
  "training": [
    
    {
      "training": 1,
      "name": "finsim-ppo-run-007",
      "description": "PPO training run for FinSim baseline.",
      "environment_name": "Demo-FinSim-Env-aks-Demo",
      "agent_id": "",
      "scenario_id": "",
      "verifier_ids": [],
      "algorithm": "PPO",
      "max_steps": 200,
      "episodes": 500,
      "reward_fn": {
        "path": "rewards/finsim_reward.py"
      },
      "simulation": {
        "speed": 1.5,
        "seed": 42,
        "episode_settings": {
          "num_episodes": 500,
          "max_steps": 200
        }
      },
      "config": {
        "learning_rate": 0.0003,
        "batch_size": 128,
        "checkpoint_interval": 100
      }
    },
    {
      "training": 2,
      "name": "finsim-ppo-run-008",
      "description": "PPO training run for FinSim baseline.",
      "environment_name": "Demo-FinSim-Env-aks-Demo",
      "agent_id": "",
      "scenario_id": "",
      "verifier_ids": [],
      "algorithm": "PPO",
      "max_steps": 200,
      "episodes": 500,
      "reward_fn": {
        "path": "rewards/finsim_reward.py"
      },
      "simulation": {
        "speed": 1.5,
        "seed": 43,
        "episode_settings": {
          "num_episodes": 500,
          "max_steps": 200
        }
      },
      "config": {
        "learning_rate": 0.0003,
        "batch_size": 128,
        "checkpoint_interval": 100
      }
    }
  ]
}

Examples

Clone or download the examples/ folder and run them in order:

File	What it shows
examples/quickstart.py	Connect, ping, list environments and tools
examples/create_environment.py	Create env → verifier → scenarios → suites
examples/verifiers.py	All 4 verifier types (rule, trajectory, LLM, composite)
examples/train_agent.py	Launch GRPO training, monitor progress, get best checkpoint
examples/evaluate_agent.py	Run evaluation, compare rollouts, export audit report

# Install the SDK
pip install rleaas

# Run examples against a local server
python examples/quickstart.py
python examples/create_environment.py
python examples/verifiers.py
python examples/train_agent.py
python examples/evaluate_agent.py

Async support

async with rleaas.AsyncClient() as client:
    status = await client.ping()

License

MIT