Filtered from nvidia/Nemotron-Math-v2 — only the hardest problems with tool-use trajectories.
Filters applied
Source: only high reasoning effort splits
changed_answer_to_majority == False
metadata.reason_high_with_tool.accuracy <= 0.375 (model gets <= 3/8 correct)
tools field non-empty (Python TIR tool definition present)
messages contain actual tool_calls (model used the tool)
expected_answer is non-null
Splits (by pass… See the full description on the dataset page: https://huggingface.co/datasets/akh-mysterio/nemotron-math-v2-hard-tool.
