Nemotron Sft Arc Agi V1: Multi-Turn Agentic Reasoning Traces for Visual Puzzles

Nemotron-SFT-ARC-AGI-v1 is a supervised fine-tuning dataset of multi-turn agentic reasoning traces. It was created by NVIDIA using nine open-weight large language models attempting to solve ARC-AGI visual-reasoning puzzles. The dataset was last updated on June 4, 2026.

Use Cases

• Training language models for multi-step reasoning based on the described agentic reasoning traces.
• Benchmarking model performance on ARC-AGI visual puzzles based on the described puzzle format.
• Studying agentic behavior and failure modes in LLMs based on the described reasoning traces.
• Developing instruction-following or chain-of-thought capabilities based on the described supervised fine-tuning context.

Strengths

• Dataset is produced by NVIDIA, a major AI research organization.
• Contains traces from nine different open-weight large language models, suggesting model diversity.
• Focuses on ARC-AGI, a known benchmark for abstract visual reasoning.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

NVIDIA

Collection Method

Traces produced by open-weight large language models attempting to solve ARC-AGI puzzles.

Freshness

Last updated 2026-06-04 04:32:57; freshness should be verified.