xrtm-train 0.3.0

The Learning/Optimization layer for XRTM.

xrtm-train

The Optimization Layer for XRTM.

xrtm-train is the engine that closes the loop. It simulates history by replaying forecast workflows against past ground-truth snapshots stored in xrtm-data, scoring the resulting forecast results with xrtm-eval, and optimizing their reasoning parameters.

Part of the XRTM Ecosystem

Layer 4: xrtm-train    → (imports all) ← YOU ARE HERE
Layer 3: xrtm-forecast → (imports eval, data)
Layer 2: xrtm-eval     → (imports data)
Layer 1: xrtm-data     → (zero dependencies)

xrtm-train sits at the top of the stack and can import from ALL other packages. Installing xrtm-train gives you the full XRTM stack.

Installation

pip install xrtm-train

This automatically installs xrtm-forecast, xrtm-eval, and xrtm-data.

Core Primitives

The Simulation Loop

The Backtester orchestrates the simulation. It ensures strict temporal isolation—agents are never exposed to data from the future.

from xrtm.train import Backtester

# Initialize components
backtester = Backtester(agent=my_agent, evaluator=my_evaluator)

# Run simulation
results = await backtester.run(dataset=historical_questions)
print(f"Mean Brier Score: {results.mean_score}")

Examples (v0.1.2+)

With the v0.6.0 architecture split, calibration and replay examples now live here:

• Calibration Demo: Adjusting confidence intervals to match reality.
• Trace Replay: Re-running a saved execution trace for debugging.
• Evaluation Harness: End-to-end backtest with metrics.

Project Structure

src/xrtm/train/
├── core/            # Interfaces & Schemas
│   └── eval/            # Calibration (PlattScaler, BetaScaler)
├── kit/             # Training utilities
│   ├── memory/          # Replay buffers
│   └── optimization/    # Training strategies
├── simulation/      # Backtester, TraceReplayer
└── providers/       # Remote training backends (future)

Development

Prerequisites:

• uv

# Install dependencies
uv sync

# Run tests
uv run pytest