Nemotron-RL-ARC-AGI-v1

Dataset Description: Nemotron-RL-ARC-AGI-v1 is a reinforcement-learning (RL) gym environment dataset of single-step ARC-AGI puzzle prompts intended for RL post-training of large language models. Each row corresponds to one ARC puzzle (a set of (input grid, output grid) demonstration pairs plus a single test input grid) rendered as a text prompt; reward is binary (1.0 / 0.0) determined by exact-match comparison against the ground-truth output grid. No LLM judge is used, no partial credit is given, and verification is fully deterministic. The dataset ships two task variants, released as Hugging Face subsets so they can be selected independently: **transductive** — the model must output the solution grid directly in `\boxed{}` format; the parsed grid is exact-matched against `expected_output`. **python_inductive** — the model must output a Python `transform(grid)` function that is executed (30-second timeout) on the test input; the resulting grid is exact-matched against `expected_output`. Both variants are single-step (one model response per problem). The same 10,000 training problems and 514 validation problems are used across both variants; only the rendered prompt and the reward extraction logic differ. The 514 validation problems are the union of the ARC-AGI-1 evaluation set (400 problems) and the ARC-AGI-2 evaluation set (120 problems, deduplicated), which are blacklisted from training to prevent leakage. Seed puzzles come from the same three publicly available sources used by the companion SFT dataset: ARC-AGI-2 (Apache-2.0), the NVARC Augmented Puzzles set published by the NVIDIA Kaggle team, and the `arc-dataset-collection` (20 community-collected subsets, predominantly under MIT or Apache-2.0 licenses). 10,000 problems are sampled uniformly at random from the ~66K eligible problems (filtered to ≤5 training examples per problem and excluding all validation problem IDs). No augmentations are released — one row per problem. The released dataset contains text pro