nemo-gym-pivot-datasets

What problem does it solve?

Converts heterogeneous agent rollout artifacts into consistent Nemo Gym pivot datasets for single-step local RL or evaluation, while validating that each expected action matches the required Responses-style row contract.

Core Features & Use Cases

• Pivot JSONL row generation: Creates one pivot row per selected assistant decision point by reconstructing the semantic pieces required by Gym’s Responses-style format.
• Gym config alignment: Produces or updates the matching Gym YAML so the pivot JSONL can be used immediately for train/eval.
• Contract validation and provenance-aware auditing: Validates JSON shape, expected_action schema, agent_ref alignment, and tool/arguments requirements, producing useful metrics to inspect skipped rows and coverage.

Quick Start

Use the nemo-gym-pivot-datasets skill to convert your rollout JSONL or trajectory artifacts into a Nemo Gym-compatible pivot JSONL and verify it with validate_pivot_dataset.py.