← Back to the article Data viewer

Data viewer

Nemotron corpus distribution vs the balanced sample — The source corpus vs the first-round sample: equal share per category.

Sample category shares vs the 89-task eval anchor — The sample vs the eval anchor. Six eval categories do not exist in the corpus.

The held-out test set, straight from each task's task.toml. Click a row for the instruction.

Tasks

#	Task	Category	Difficulty	Expert min	Tags

No tasks match.

Rows

#	Category	Level	Turns	KChars	Source task	Task

No rows match.

Generated by scripts/build_data_viewer.py in the training repo. Re-run after data changes; numbers come from the jsonl and the task.toml files, never hand-edited.