Nemotron-RL-ARC-AGI-v1: Single-Step ARC Puzzle Prompts for RL Post-Training

A reinforcement-learning gym environment dataset of single-step ARC-AGI puzzle prompts for post-training large language models. Each row contains a text prompt rendering of an ARC puzzle with input-output grid pairs and a test input, with binary reward determined by exact-match comparison. The dataset was created by NVIDIA and last updated on 2026-06-04.

Use Cases

• Reinforcement learning post-training of LLMs based on the described binary reward structure.
• Evaluating LLM performance on abstract reasoning tasks based on the ARC puzzle format.
• Benchmarking agent performance in single-step environments based on the grid-based puzzle prompts.

Strengths

• Reward is defined by a clear, objective metric (binary exact-match comparison).
• Designed for a specific, advanced use case (RL post-training of LLMs).
• Source is a reputable organization (NVIDIA).

Limitations

• Row count, column definitions, and file formats are unknown, limiting suitability assessment.
• Column-level documentation is absent; field semantics must be inferred after download.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

NVIDIA

Collection Method

Likely generated or curated for reinforcement learning research.

Time Range

null

Freshness

Last updated 2026-06-04 05:08:56; freshness should be verified.

Geography

null