verifiablelabs/

prm-dataset-v0.0.1

Per-step process-reward training data for the Verifiable Labs SDK, produced by the Phase 30 process-reward pipeline.

Traces: 120 Total steps: 120 Traces with per-step frontier judgments: 28 Frontier judge: anthropic/claude-sonnet-4 (when judged) Source mix: env: 92 judgment: 28

Each row is a JSON object with the following fields:

field type meaning

row_id str unique id

env_id str env… See the full description on the dataset page: https://huggingface.co/datasets/verifiablelabs/prm-dataset-v0.0.1.