RL datasets live Contribute · earn $PETRI

Crowdsource RL datasets for science.

Petri pays contributors in $PETRI to build verified RL training data — observation, action, reward rows labs can export and train on

OpenLaband pick an open lab request

RL

dataset format

3

verified data types

1.0

reward per valid row

JSONL

export for labs

RL-ready science data.

Every submission becomes a training row — observation, action, reward, ground truth — exported as JSONL for reinforcement learning and supervised fine-tuning.

01

Culture photos

Real mold, plates, plant growth — blur-checked, tagged, deduplicated.

02

Lab notes OCR

Handwritten or printed pages — must be sharp with visible text/lines.

03

Verified voice

Read a science sentence — speech must match transcript before payout.

jsonl · observation/action/rewardRL dataset

Culture photo

Vision rows with labels — reward on verified upload.

Lab notes

OCR pairs — image observation, text ground truth.

Voice read

Speech rows — audio action, verified transcript reward.

Better data, better scores.

Models trained on Petri contributor data jump on held-out science benchmarks — genes, stats pipelines, lab photos, and more.

Pass rate by benchmark

Baseline vs after Petri RL training data

BaselineAfter Petri

Pass rate (%) · held-out tasks · n=500 per benchmark

$PETRI

Paid to build RL datasets

Labs need verifiable reward signals, not random uploads. $PETRI pays contributors when their row lands in the export — ready for RL and fine-tuning pipelines.

Contribute data

Photos, notes, voice — quality-checked

RL row minted

observation · action · reward · ground_truth

Earn $PETRI

Paid when it hits the dataset