Open Agentrl Grpo 2K: A Compact Dataset for GRPO Training Across Math, Science, and Code

Open-AgentRL GRPO 2K is a compact dataset containing approximately 2,000 samples for GRPO training. It was created by y-ohtani and last updated on February 28, 2026. The dataset is constructed by balanced sampling from five sources: DeepScaleR-Preview (374 math items), NuminaMath-1.5 (359 math items), Omni-MATH (366 math items), GPQA Diamond (198 science items), and LeetCodeDataset (351 code items).

Use Cases

• Training GRPO agents for mathematical problem-solving based on the math domain content.
• Fine-tuning language models on scientific Q&A for physics, chemistry, and biology based on the GPQA science subset.
• Developing code generation agents using programming problems based on the LeetCode code samples.
• Benchmarking agent performance across multiple reasoning domains based on the balanced sampling from five distinct sources.

Strengths

• Contains a balanced mix of approximately 2,000 samples from five distinct data sources.
• Includes 374 math problems from DeepScaleR-Preview, 359 from NuminaMath-1.5, and 366 from Omni-MATH.
• Adds 198 science questions from the GPQA Diamond benchmark covering physics, chemistry, and biology.
• Incorporates 351 programming problems from a LeetCode dataset.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• The dataset is described as 'compact'; its scale may be limited for large-scale model training.

Provenance

Source

Aggregated from five datasets on Hugging Face: agentica-org/DeepScaleR-Preview, AI-MO/NuminaMath-1.5, KbsdJames/Omni-MATH, Idavidrein/gpqa (gpqa_diamond), and newfacade/LeetCodeDataset.

Collection Method

Created by balanced, equal sampling from each of the five source datasets.

Time Range

null

Freshness

Last updated 2026-02-28 10:18:27; freshness should be verified.

Geography

null