dataset_info:
- config_name: default
features:
- name: id
dtype: string
- name: query
dtype: string
- name: texts
sequence: string
- name: context_spans
sequence:
sequence:
sequence: int64
- name: context_spans_relevance
sequence:
sequence: int64
- name: labels
sequence: int64
- name: teacher_scores.japanese-reranker-xsmall-v2
sequence: float64
- name: teacher_scores.japanese-reranker-base-v2
sequence: float64
- name: teacher_scores.gte-reranker-modernbert-base
sequence: float64
- name: teacher_scores.ruri-v3-reranker-310m
sequence: float64
- name: teacher_scores.bge-reranker-v2-m3
sequence: float64
splits:
- name: train
num_bytes: 2120629817
num_examples: 492729
- name: validation
num_bytes: 21576804
num_examples: 5000
- name: test
num_bytes: 21525217
num_examples: 5000
download_size: 887176985
dataset_size: 2163731838
- config_name: freq2
features:
- name: id
dtype: string
- name: query
dtype: string
- name: texts
sequence: string
- name: context_spans
sequence:
sequence:
sequence: int64
- name: context_spans_relevance
sequence:
sequence: int64
- name: labels
sequence: int64
- name: teacher_scores.japanese-reranker-xsmall-v2
sequence: float64
- name: teacher_scores.japanese-reranker-base-v2
sequence: float64
- name: teacher_scores.gte-reranker-modernbert-base
sequence: float64
- name: teacher_scores.ruri-v3-reranker-310m
sequence: float64
- name: teacher_scores.bge-reranker-v2-m3
sequence: float64
splits:
- name: train
num_bytes: 1112849819
num_examples: 260436
- name: validation
num_bytes: 4329211
num_examples: 1000
- name: test
num_bytes: 4297213
num_examples: 1000
download_size: 469610114
dataset_size: 1121476243
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: validation
path: data/validation-*
- split: test
path: data/test-*
- config_name: freq2
data_files:
- split: train
path: freq2/train-*
- split: validation
path: freq2/validation-*
- split: test
path: freq2/test-*
This dataset includes extracted hard negatives, along with relevance scores between questions and text spans, and reranker scores. It can be used for training models such as OpenProvence.
This dataset is based on MS MARCO and follows the license of the original MS MARCO dataset.
Available Subsets
Recommended: freq2 — 260,436 train / 1,000 validation / 1,000 test rows.
This split applies the MD5-based frequency filter with N=2, removing any row whose passages have already surfaced more than twice. The trimmed validation/test sets are freshly sampled (1k each) so their duplication profile matches the training subset (≈10% duplicate texts). Use this configuration when you want balanced coverage with significantly reduced redundancy and faster training cycles.
default — 492,729 train / 5,000 validation / 5,000 test rows.
Full MS MARCO conversion after attaching teacher scores (gte-reranker-modernbert-base, ruri-v3-reranker-310m, bge-reranker-v2-m3, japanese-reranker-*). Intra-row duplication is minimal (<0.01%), so the split keeps the complete breadth of the corpus. Reach for this if you need the original distribution or larger evaluation sets.