Launch Proba Post-training, engineered.
Proba is an alignment platform that turns post-training from trial and error into engineering.
Capabilities
Predict training outcomes, diagnose model behavior, debug failures, and steer what your model learns.
Built for
Anyone post-training open-weight models: ML engineers, startups, researchers, hobbyists, and their agents.
Instruments
Sparse autoencoders, linear probes, and other interpretability tooling from validated open research.
Put Proba to work.
Cheap answers to expensive questions.
Know what a dataset will teach.
Proba analyzes your preference dataset through your model's internals. You get a readout of the behaviors it will shift, which way, and by how much.
Compare datasets before you commit.
A readout takes minutes, not a training run. Compare candidate datasets side by side and commit to one before you spend any compute.
Fix the data, not the checkpoint.
When a readout shows data pulling the wrong way, Proba edits it, doubles down on what works, and re-checks with another prediction. No retraining to find out.
Hand the loop to an agent.
Every readout is a report you can read and a document a machine can, behind one API call. An agent can run the whole loop while you review the trail.
See inside the model.
-
Step 01
Bring your data.
Upload your post-training dataset to Proba.
-
Step 02
Interpret through the model.
Proba runs your data through the model and reads the activations with interpretability instruments.
-
Step 03
Proba creates a report.
A readout of what the data will teach: which behaviors shift and how far, with high-risk findings flagged first.
-
Step 04
You iterate on the dataset.
Edit, re-check, repeat. Then train once on the engineered dataset.
Proba runs your data through the model and reads the activations with mechanistic interpretability instruments.
What comes back is a report you can act on.
Humans read it. Agents run it.
Proba is built to be driven by agents as well. Point yours at the MCP skill and API reference, and it can run the predict, edit, verify loop on its own.
For agents
Explore Aurelius
Subnet
