FROM THE RECURVIA RL LAB

See what your training is actually doing.

We're a deep reinforcement learning lab that built Metrana to log our RL and foundation-model training at full rate: every detail. And a bill that can't surprise you. Now open to partner labs.

Book a Call Read the Benchmark

python

import metrana

metrana.init(
    workspace_name="my-workspace",
    project_name="my-project",
    run_name="quickstart-ml",
    config={"optimizer": "adam", "lr": 3e-4, "batch_size": 256},   # logged as run attributes
)

Metrana plugs into your training workflow with just a few lines of code - Start capturing system-wide signals immediately.

Copied to clipboard

Under-logging hides the exact step that failed.

Every missing metric is a blind spot in your debugging process. Metrana captures the complete execution history, so failures can be traced to their true point of origin instead of guessed from incomplete data.

The dashboard said fine. The run wasn't.

Mean reward climbs while the policy quietly hacks the reward; loss and reward converge while some environments diverge. Aggregates can lead to false confidence, whereas environment level metrics give deeper understanding. Most tools only make the first one easy.

Your data is hierarchical. Your tracker isn’t.

RL experiments contain environments, episodes, and timesteps. Foundation models contain layers, experts, heads, and tokens. Metrana preserves that structure, so you can navigate your data the way your systems are actually built.

The metric was there. It wasn't found fast enough.

Thousands of series, one dashboard. Finding the one that moved before the collapse means digging manually - by the time you have, the run's already burned the compute.

We hit all these problems on our own runs. That's why we built Metrana.

Ingest everything

Full-rate logging of thousands of signals per run. No sampling, no dropping. Sustained 85k series entries and 5M+ data points per second.

RL-native, not RL-only

Environments, episodes, and steps are first-class citizens, not folder names. Watch what happened, not just how the curve moved.

Transparent pricing.

Pay for what you log. Set a spending limit; we'll alert you then slow you down instead of charging past it.

Analysis agent (early access).

Explains and recommends today. Automates soon.

For reinforcement-learning teams

Chips, plants, data centres, robotics, post-training LLM-based agents; workloads where a run is thousands of parallel environments, not a single execution.

Per-environment, per-episode, per-step views at full logging rate - no more downsampling valuable signals.
Replay any environment, any episode, at any step with synchronized video rendering and metrics for true behavioural debugging.
Fork your experiments at both the run and environment level, then compare across forks.

Read the RL guide

For foundation-model teams

Foundation model pre-training and supervised fine-tuning isn’t flat either: epochs, modules, layers, experts, heads; Metrana lets you organise, inspect and compare them all natively, so you can debug behaviour at the level where it actually emerges.

Metric volumes that break default tooling: Log every signal at full resolution with no forced downsampling hiding.
Organise experiments into natural hierarchies: modules, layers, experts, heads, losses - whatever structure best reflects your model and training setup.
Fork experiments and compare checkpoints from the exact moment they diverged.

Talk to us about post-training

Copied to clipboard

Built by a lab. Run as a partnership.

Metrana comes out of Recurvia (recurvia.ai), a deep reinforcement-learning lab. We built this instrumentation because our own training runs outgrew every tracker we tried - and we weren't willing to log less or slow our workflows to fit the tool.

Rather than keep it internal, we've opened Metrana for select partner labs: you get direct roadmap input, early access to new capabilities, and engineering support from the team building it.

Become a partner lab About Recurvia

Sustained ingestion of 85,000 series entries and 5M+ data points per second; full specs, methodology, and hardware in the write-up.

Read the benchmark