Proba

Predict what your fine-tuning data will do, before you train

Launch Proba

Applications

Predict training outcomes, diagnose model behavior, debug failures, and steer what your model learns.

Instruments

Sparse autoencoders, linear probes, and other interpretability tooling from validated open research, curated and calibrated for reliable readouts.

Built for

Anyone post-training open-weight models: ML engineers, startups, researchers, hobbyists, and the agents working on their behalf.

See inside the model

A platform for reading what's happening inside a model. No trial runs, no guesswork. Proba is predictive data debugging: read what a dataset will teach, and catch contamination or mimicry, before a single training step.

The Proba pipeline in three steps: a sparse autoencoder reads preference pairs to show what your data will teach, the link from data clusters to concept clusters like increased sycophancy is traced, and the debug step intervenes: filter the data or reshape the reward

Put Proba to work

Cheap answers to expensive questions.

Predict

See what your data will teach.

What a dataset teaches depends on the model learning it. Proba reads your data, DPO or KTO preference pairs, through your model's internals: each concept it will affect, the direction and magnitude, and a confidence score on every readout.

Compare

Rank candidate datasets.

Five candidate datasets used to mean five training runs and an eval suite over the checkpoints. Proba predicts with batched forward passes instead of compute-heavy training runs: minutes per candidate, so you compare all five and commit to one before any real spend.

Projected concept shifts · Δ chosen − rejected

Feature	Dataset A	Dataset B	Dataset C
Direct instruction-following	+0.42	+0.19	−0.08
Evaluation awareness	+0.29	+0.04	+0.33
Verbose elaboration	+0.18	−0.14	+0.02
Sycophantic agreement	−0.15	+0.21	−0.24
Refusal boundary	−0.09	+0.12	+0.02
⋮	⋮	⋮	⋮
Structured formatting	+0.03	+0.22	−0.10

Compare how each dataset moves the features you care about

Intervene

Fix the data, not the checkpoint.

Armed with a readout of what a dataset will teach, you can intervene to improve your results. Edit data that pushes the wrong way, double down on what works, and re-verify with another prediction, all before any compute is spent on training.

Trace the shift to the data responsible, fix it, and confirm the fix

Iterate

Humans read it. Agents run it.

Every readout arrives as an interactive web report with a matching canonical JSON document. That means the predict, edit, verify loop doesn't need you in it: every readout is a single API call, so an agent can request readouts, edit data, and verify fixes in the background while you review the trail.

Agents connect via API and iterate using Proba before committing to training

How it works

Built on open interpretability research.

Proba runs your data through the model, reading activations via mechanistic interpretability instruments: SAEs, linear probes, logits, and more. These tools rely on forward passes over the data, which is why readouts cost a fraction of the time and compute of a training run.

Sparse autoencoders (SAE)

Linear probes

Logit readouts

Lenses

Validated instruments from open interpretability research

Try Proba today

Select your model, post-training method, and dataset - get a readout in minutes.