Target-Oriented Pretraining Data Selection via Neuron-Activated Graph

Create an open-source implementation of Neuron-Activated Graph Ranking for data selection. This allows devs to curate high-quality pretraining data without the black-box opacity of traditional embedding-based similarity.

Suggested repo: nag-rank

"Data selection, unboxed: Use neuron activations to curate better models."

Estimated effort: 50h