metadata
dataset_info:
features:
- name: query
dtype: string
- name: hits
list:
- name: content
dtype: string
- name: docid
dtype: string
- name: qid
dtype: int64
- name: rank
dtype: int64
- name: score
dtype: float64
splits:
- name: train
num_bytes: 154379410
num_examples: 21100
download_size: 5482226
dataset_size: 154379410
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
This is reranking dataset built from 179 queris from MSMARCO-V2 passages set.
Below are the data creation process:
# Multiple Random Sampling
import copy
import random
replicate_rank_results = []
replicate_times = 100
for item in combined_rank_results:
for _ in range(replicate_times):
new_item = copy.deepcopy(item)
# Randomly select 20 hits from original hits list
random_select_index = random.sample(range(len(item['hits'])), min(20, len(item['hits'])))
random_select_index.sort()
new_item['hits'] = [new_item['hits'][i] for i in random_select_index]
replicate_rank_results.append(new_item)
print(len(replicate_rank_results))