Hugging Face's logo

Datasets:
hotchpotch
/
japanese-splade-v1-hard-negatives

Modalities:
Text
Formats:
parquet
Size:
10M - 100M
Libraries:
Datasets
Dask
Dataset card Viewer Files Community
1
japanese-splade-v1-hard-negatives / README.md
hotchpotch's picture
hotchpotch
Update README.md
4d3f913 verified
preview code
|
raw
history blame contribute delete
6.68 kB
---
dataset_info:
- config_name: mmarco-collection
features:
- name: text
dtype: string
- name: original_row_id
dtype: int64
splits:
- name: train
num_bytes: 3884756138
num_examples: 8829813
download_size: 2267388203
dataset_size: 3884756138
- config_name: mmarco-dataset
features:
- name: original_row_id
dtype: int64
- name: anc
dtype: string
- name: pos_ids
sequence: int64
- name: neg_ids.original
sequence: int64
- name: neg_ids.japanese-splade-base-v1-mmarco-only.top100
sequence: int64
- name: neg_sims.japanese-splade-base-v1-mmarco-only.top100
sequence: float64
- name: neg_ids.japanese-splade-base-v1-mmarco-only.other100
sequence: int64
- name: neg_sims.japanese-splade-base-v1-mmarco-only.other100
sequence: float64
- name: score.bge-reranker-v2-m3.pos_ids
sequence: float64
- name: score.bge-reranker-v2-m3.neg_ids.japanese-splade-base-v1-mmarco-only.top100
sequence: float64
- name: score.bge-reranker-v2-m3.neg_ids.japanese-splade-base-v1-mmarco-only.other100
sequence: float64
- name: neg_ids.japanese-splade-base-v1_5.top100
sequence: int64
- name: neg_sims.japanese-splade-base-v1_5.top100
sequence: float64
- name: neg_ids.japanese-splade-base-v1_5.other100
sequence: int64
- name: neg_sims.japanese-splade-base-v1_5.other100
sequence: float64
- name: score.bge-reranker-v2-m3.neg_ids.japanese-splade-base-v1_5.top100
sequence: float64
- name: score.bge-reranker-v2-m3.neg_ids.japanese-splade-base-v1_5.other100
sequence: float64
splits:
- name: train
num_bytes: 4123278189
num_examples: 391060
download_size: 3397702699
dataset_size: 4123278189
- config_name: mqa-collection
features:
- name: text
dtype: string
- name: original_row_id
dtype: int64
splits:
- name: train
num_bytes: 5499685825
num_examples: 11852254
download_size: 3335944299
dataset_size: 5499685825
- config_name: mqa-dataset
features:
- name: original_row_id
dtype: int64
- name: anc
dtype: string
- name: pos_ids
sequence: int64
- name: neg_ids
sequence: 'null'
- name: neg_ids.japanese-splade-base-v1-mmarco-only.top100
sequence: int64
- name: neg_sims.japanese-splade-base-v1-mmarco-only.top100
sequence: float64
- name: neg_ids.japanese-splade-base-v1-mmarco-only.other100
sequence: int64
- name: neg_sims.japanese-splade-base-v1-mmarco-only.other100
sequence: float64
- name: score.bge-reranker-v2-m3.pos_ids
sequence: float64
- name: score.bge-reranker-v2-m3.neg_ids.japanese-splade-base-v1-mmarco-only.top100
sequence: float64
- name: score.bge-reranker-v2-m3.neg_ids.japanese-splade-base-v1-mmarco-only.other100
sequence: float64
- name: neg_ids.japanese-splade-base-v1_5.top100
sequence: int64
- name: neg_sims.japanese-splade-base-v1_5.top100
sequence: float64
- name: neg_ids.japanese-splade-base-v1_5.other100
sequence: int64
- name: neg_sims.japanese-splade-base-v1_5.other100
sequence: float64
- name: score.bge-reranker-v2-m3.neg_ids.japanese-splade-base-v1_5.top100
sequence: float64
- name: score.bge-reranker-v2-m3.neg_ids.japanese-splade-base-v1_5.other100
sequence: float64
splits:
- name: train
num_bytes: 4189340797
num_examples: 427532
download_size: 3448383271
dataset_size: 4189340797
- config_name: msmarco-ja-collection
features:
- name: text
dtype: string
- name: original_row_id
dtype: int64
- name: text_en
dtype: string
splits:
- name: train
num_bytes: 6761932587
num_examples: 8841823
download_size: 3637581470
dataset_size: 6761932587
- config_name: msmarco-ja-dataset
features:
- name: original_row_id
dtype: int64
- name: anc
dtype: string
- name: pos_ids
sequence: int64
- name: neg_ids.japanese-splade-base-v1_5.top100
sequence: int64
- name: neg_sims.japanese-splade-base-v1_5.top100
sequence: float64
- name: neg_ids.japanese-splade-base-v1_5.other100
sequence: int64
- name: neg_sims.japanese-splade-base-v1_5.other100
sequence: float64
- name: score.bge-reranker-v2-m3.pos_ids
sequence: float64
- name: score.bge-reranker-v2-m3.neg_ids.japanese-splade-base-v1_5.top100
sequence: float64
- name: score.bge-reranker-v2-m3.neg_ids.japanese-splade-base-v1_5.other100
sequence: float64
- name: neg_ids.japanese-splade-base-v1-mmarco-only.top100
sequence: int64
- name: neg_sims.japanese-splade-base-v1-mmarco-only.top100
sequence: float64
- name: neg_ids.japanese-splade-base-v1-mmarco-only.other100
sequence: int64
- name: neg_sims.japanese-splade-base-v1-mmarco-only.other100
sequence: float64
- name: score.bge-reranker-v2-m3.neg_ids.japanese-splade-base-v1-mmarco-only.top100
sequence: float64
- name: score.bge-reranker-v2-m3.neg_ids.japanese-splade-base-v1-mmarco-only.other100
sequence: float64
- name: anc_en
dtype: string
splits:
- name: train
num_bytes: 4911934526
num_examples: 502931
download_size: 4025308774
dataset_size: 4911934526
configs:
- config_name: mmarco-collection
data_files:
- split: train
path: mmarco-collection/train-*
- config_name: mmarco-dataset
data_files:
- split: train
path: mmarco-dataset/train-*
- config_name: mqa-collection
data_files:
- split: train
path: mqa-collection/train-*
- config_name: mqa-dataset
data_files:
- split: train
path: mqa-dataset/train-*
- config_name: msmarco-ja-collection
data_files:
- split: train
path: msmarco-ja-collection/train-*
- config_name: msmarco-ja-dataset
data_files:
- split: train
path: msmarco-ja-dataset/train-*
---
[日本語 SPLADE v2](https://huggingface.co/hotchpotch/japanese-splade-v2) の学習に用いたデータセットです。
SPLADE モデルである [hotchpotch/japanese-splade-base-v1-mmarco-only](https://huggingface.co/hotchpotch/japanese-splade-base-v1-mmarco-only), [japanese-splade-base-v1_5](https://huggingface.co/hotchpotch/japanese-splade-base-v1_5) を用いてハードネガティブマイニングを行なっています。また BAAI/bge-reranker-v2-m3 を用いたリランカースコアを付与しています。
- mqa, mmarco は[hpprc/emb](https://huggingface.co/datasets/hpprc/emb) のデータを用いています。
- mqa の query 作成には MinHash を用い約40万件になるようフィルタしました。
- msmarco-ja は [hpprc/msmarco-ja](https://huggingface.co/datasets/hpprc/msmarco-ja)のデータを用いています。
ライセンスは、各データセットのライセンスを継承します。