Datasetpublic

Nithish2410/msm-fineweb-hiconf-top10-mix75-25-grpo

MS MARCO + FineWeb high-confidence top10 75/25 GRPO mix Purpose: conservative GRPO ablation after pure FineWeb filtered training still reduced NQ. Sources: MS MARCO querygen: Nithish2410/msm-q3-re-ranked-t100-querygen-diverse FineWeb filtered querygen: Nithish2410/fineweb-all-domain-e5v2qwen-task-10k-passages-hiconf-top10 Rows: total: 12000 MS MARCO rows: 9000 (75%, cycled from the 2028-row source with unique row ids) FineWeb rows: 3000 (25%, sampled without replacement with… See the full description on the dataset page: https://huggingface.co/datasets/Nithish2410/msm-fineweb-hiconf-top10-mix75-25-grpo.

sourceHugging Faceupdated 2mo agoView on Hugging Face

0likes8downloads

Dataset Card

MS MARCO + FineWeb high-confidence top10 75/25 GRPO mix

Purpose: conservative GRPO ablation after pure FineWeb filtered training still reduced NQ.

Sources:

—MS MARCO querygen: Nithish2410/msm-q3-re-ranked-t100-querygen-diverse
—FineWeb filtered querygen: Nithish2410/fineweb-all-domain-e5v2qwen-task-10k-passages-hiconf-top10

Rows:

—total: 12000
—MS MARCO rows: 9000 (75%, cycled from the 2028-row source with unique row ids)
—FineWeb rows: 3000 (25%, sampled without replacement with seed 100)

This dataset contains only the query split. The intended training command uses existing ES indices with --skip-indexing --elastic-index msmarco_train,fineweb.