Datasets:
metadata
dataset_info:
features:
- name: question
dtype: string
- name: answer
dtype: string
- name: articles
list: string
- name: instruction_full
dtype: string
- name: metadata
list: string
- name: instruction_proxy
dtype: string
splits:
- name: train
num_bytes: 1166992621
num_examples: 4136
- name: val
num_bytes: 113053705
num_examples: 400
- name: test
num_bytes: 169391827
num_examples: 600
download_size: 1446932927
dataset_size: 1449438153
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: val
path: data/val-*
- split: test
path: data/test-*
license: apache-2.0
language:
- en
tags:
- long-context
pretty_name: proxycot-hotpotqa
size_categories:
- 1K<n<10K
This is the HotpotQA data that we used in our ProxyCoT project (https://aclanthology.org/2026.acl-long.1917/), and it is based on long-context reasoning (32K-128K tokens).
HotpotQA here is a new version originally from https://aclanthology.org/2026.acl-long.1917/ with extended contexts. For more details on the context extension, refer to the ProxyCoT paper.
To use our dataset, please follow the code below.
train_samples = load_dataset("oaimli/proxycot-hotpotqa", split="train")
dev_samples = load_dataset("oaimli/proxycot-hotpotqa", split="val")
test_samples = load_dataset("oaimli/proxycot-hotpotqa", split="test")
for sample in train_samples:
question = sample["question"]
answer = sample["answer"]
metadata = sample["metadata"]
articles = sample["articles"]
instruction_proxy = sample["instruction_proxy"]
instruction_full = sample["instruction_full"]
instruction_proxy = instruction_proxy.replace("<question>", question)
instruction_proxy = instruction_proxy.replace("<articles>", "\n\n\n".join(metadata))
conversation_proxy = [{"role": "user", "content": instruction_proxy}]
instruction_full = instruction_full.replace("<question>", question)
instruction_full = instruction_full.replace("<articles>", "\n\n\n".join(articles))
conversation_full = [{"role": "user", "content": instruction_full}]