tviz Documentation

tviz is a local visualization dashboard for RL training runs. Think TensorBoard, but designed for Tinker and reinforcement learning workflows.

Installation

pip install tviz tinker tinker-cookbook

git clone https://github.com/sdan/tviz.git cd tviz && bun install

Guides

Getting Started

Fine-tune Llama-3.1-8B on GSM8K math problems with tviz logging. Based on the tinker-cookbook RL basic recipe.

Examples

Ready-to-run examples: GSM8K RL, supervised fine-tuning, custom loops.

API Reference

Complete TvizLogger API: log_metrics, log_hparams, log_rollout, and more.

Tinker Integration

Add tviz to tinker-cookbook's MultiplexLogger for seamless visualization.

Quick Example

Add tviz to any Tinker training script with just a few lines:

from tviz import TvizLogger

# Create logger
logger = TvizLogger(run_name="my-training-run")
logger.log_hparams({"model": "llama-8b", "lr": 4e-5})

# In your training loop
for step in range(num_steps):
    # ... training code ...
    logger.log_metrics({"loss": loss, "reward": reward}, step=step)

# Cleanup
logger.close()

View Dashboard

cd tviz && bun dev

Open http://localhost:3003 to view your training runs.

Data storage: By default, tviz stores data in ~/.tviz/tviz.db. Override with TVIZ_DB_PATH.