Datasets:
metadata
dataset_info:
features:
- name: idx
dtype: int64
- name: names
dtype: string
- name: parallel_chain
dtype: string
- name: parallel_total_val
dtype: float64
- name: parallel_lastname
dtype: string
- name: parallel_single_val
dtype: float64
- name: forward_chain
dtype: string
- name: forward_total_val
dtype: float64
- name: forward_lastname
dtype: string
- name: forward_single_val
dtype: float64
- name: backward_chain
dtype: string
- name: backward_total_val
dtype: float64
- name: backward_lastname
dtype: string
- name: backward_single_val
dtype: float64
- name: chaotic_chain
dtype: string
- name: chaotic_total_val
dtype: float64
- name: chaotic_lastname
dtype: string
- name: chaotic_single_val
dtype: float64
splits:
- name: k5
num_bytes: 178184
num_examples: 200
- name: k10
num_bytes: 333938
num_examples: 200
- name: k20
num_bytes: 647136
num_examples: 200
- name: k50
num_bytes: 1582289
num_examples: 200
- name: k100
num_bytes: 3142590
num_examples: 200
- name: k200
num_bytes: 6266799
num_examples: 200
download_size: 4072876
dataset_size: 12150936
configs:
- config_name: default
data_files:
- split: k5
path: data/k5-*
- split: k10
path: data/k10-*
- split: k20
path: data/k20-*
- split: k50
path: data/k50-*
- split: k100
path: data/k100-*
- split: k200
path: data/k200-*
task_categories:
- question-answering
language:
- en
tags:
- llm-evaluation
- long-context
- reasoning
- benchmark
NeedleChain: Measuring Intact Long-Context Reasoning Capability of Large Language Models
Github: Official github repository
Paper: Official Paper
NeedleChain is a benchmark designed to evaluate LLMs' intact long-context understanding. Every provided context consists of query-relevant information, requiring a comprehensive understanding to answer the given query.
For manual creation of NeedleChain datasets, please refer to our official github repository.