tviz 0.2.2

tviz is a visualization dashboard for the Tinker API

tviz

tviz is a local dashboard for visualizing RL training runs. Similar to wandb, you add a logger to your training loop. Metrics, rollouts, and trajectories are stored in a local SQLite database and displayed in real-time.

Installation

pip install tviz

Quick Start

from tviz import TvizLogger

logger = TvizLogger(run_name="gsm8k_rl")
logger.log_hparams({"model": "llama-3.2-1b", "lr": 1e-4})
print(f"View at: {logger.get_logger_url()}")

for step in range(100):
    # ... training code ...
    logger.log_metrics({"reward": 0.5, "loss": 0.1}, step=step)
    logger.log_rollouts(rollouts, step=step)

logger.close()

Dashboard

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

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

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

API Reference

TvizLogger

TvizLogger(
    db_path: str | None = None,    # Path to SQLite db (default: ~/.tviz/tviz.db)
    run_name: str | None = None,   # Human-readable run name
    modality: str = "text"         # "text" or "vision"
)

Methods

Method	Description
log_hparams(config)	Log hyperparameters dict
log_metrics(metrics, step)	Log scalar metrics for a step
log_rollouts(rollouts, step)	Log trajectory rollouts
get_logger_url()	Get dashboard URL for this run
close()	Mark run complete, close connection

Rollout Format

logger.log_rollouts([
    {
        "group_idx": 0,
        "prompt_text": "What is 2+2?",
        "trajectories": [
            {"trajectory_idx": 0, "reward": 1.0, "output_text": "4"},
            {"trajectory_idx": 1, "reward": 0.0, "output_text": "5"},
        ],
    }
], step=100)

Tinker Integration

from tviz.adapters.tinker import from_tinker_batch

# After rollouts:
rollouts = from_tinker_batch(trajectory_groups_P, tokenizer=tokenizer)
logger.log_rollouts(rollouts, step=i_batch)

Modalities

• text: LLM RL (math, code, chat)
• vision: VLM tasks with images and maps

License

MIT