DoYangTan/EQA_DATASET

UniEQA: A Hierarchical EEG Question-Answering Benchmark

Overview

UniEQA is a large-scale EEG question-answering benchmark designed to study generalized neural signal understanding across heterogeneous tasks, datasets, and reasoning levels. Instead of framing EEG analysis as a collection of isolated task-specific classifiers, UniEQA organizes EEG understanding into a unified QA format and a hierarchical task taxonomy.

The benchmark is built around three levels of abstraction:

• L1: direct observation of waveform patterns and basic physiological state.
• L2: translation of EEG into task semantics, stimulus attributes, or control intent.
• L3: higher-level inference and multi-attribute reasoning over EEG segments.

This repository currently provides two JSONL files:

• train.jsonl
• test.jsonl

Each line is a standalone QA example associated with an EEG segment and includes the question, answer, task metadata, and the path to the corresponding EEG sample file.

Motivation

Most prior EEG benchmarks focus on one dataset, one task family, or one prediction granularity at a time. UniEQA is intended to support a broader setting:

• A unified benchmark across multiple public EEG datasets.
• A shared QA formulation for classification, decoding, inference, and reasoning.
• A hierarchical view of EEG understanding, from low-level signal observation to high-level semantic interpretation.
• A practical testbed for multitask EEG modeling, transfer learning, and foundation-model style research.

Benchmark Highlights

• 1,490,476 QA examples across train and test.
• 14 EEG task sources spanning seizure detection, motor imagery, inner speech, emotion, neurofeedback, photic stimulation, cognitive monitoring, and clinical inference.
• 3 hierarchy levels: L1, L2, and L3.
• 6 task families ranging from waveform detection to multi-step semantic reasoning.
• 3 answer formats: open, binary_choice, and multi_choice.
• 58 unique task + question_type + task_family combinations.

Split Summary

Split	Records
train	1,392,944
test	97,532
total	1,490,476

Hierarchical Task Structure

Level	Role	Task families in this release
L1	Direct signal observation	Basic physiological state observation, Waveform and abnormal pattern detection
L2	Semantic translation from EEG to task meaning	Motor intention and control command decoding, Contextual Semantic Decoding
L3	Higher-level inference and compositional reasoning	High-level state inference, Multi-step semantic reasoning

Level Distribution

Level	Records
L1	531,162
L2	602,587
L3	356,727

Task Family Distribution

Task family	Records
Basic physiological state observation	264,202
Contextual Semantic Decoding	582,931
High-level state inference	112,017
Motor intention and control command decoding	19,656
Multi-step semantic reasoning	244,710
Waveform and abnormal pattern detection	266,960

Source Tasks

Task	Records	Representative capability
asym	570,845	Resting/task state, response-hand decoding, emotion and multi-attribute reasoning.
bciciv	2,800	Motor imagery command decoding.
chbmit	748	Seizure vs non-seizure waveform detection.
dementia	4,535	Clinical group inference from resting EEG.
exo	16,856	Gait transition decoding for exoskeleton control.
gal	13,168	Grasp-and-lift state, weight, friction, and joint reasoning.
is	2,843	Rest vs inner-speech state and imagined-word decoding.
mom	384,029	Task relevance, trial type, and combined event reasoning.
photo	1,688	Resting/photic condition and stimulation frequency decoding.
seed	75,780	Emotion-state inference.
surprise	130,808	Task variant, trial type, and joint reasoning.
tdcs	19,904	Stimulation session and motor-imagery condition reasoning.
tusz	266,212	Electrographic seizure presence and class detection.
vrnf	260	Rest vs neurofeedback state recognition.

Data Format

Each line in train.jsonl and test.jsonl is a JSON object with the following fields:

Field	Description
task	Task or dataset identifier such as asym, tusz, or mom.
subject_id	Subject identifier within the source task.
file	Relative path to the EEG segment file associated with the QA entry.
level	Hierarchy level: L1, L2, or L3.
attribute_type	The target semantic attribute for the QA example.
question_type	One of open, binary_choice, or multi_choice.
question	Natural-language question about the EEG segment.
options	Candidate answers for choice-based questions. Empty or absent for open questions.
answer	Gold answer, either as text or as an option key such as a, b, or c.
task_family	High-level capability group within the benchmark taxonomy.

Example Record

{
  "task": "asym",
  "subject_id": "sub1",
  "file": "asym/sub1/sub1_trial_00000_task.pkl",
  "level": "L3",
  "attribute_type": "state+response_hand",
  "question_type": "multi_choice",
  "question": "What characterizes this EEG segment in terms of both the experimental task and the hand used for the subject's response?",
  "options": [
    "a. This segment was recorded during a word task in which the subject responded with the left hand.",
    "b. This segment was recorded during a resting state with eyes closed, with no task-related hand response.",
    "c. This segment was recorded during an emotion task in which the subject responded with the left hand.",
    "d. This segment was recorded during a face identity task in which the subject responded with the right hand.",
    "e. This segment was recorded during a house task in which the subject responded with the right hand."
  ],
  "answer": "c",
  "task_family": "Multi-step semantic reasoning"
}

Coverage of Question Types

Question type	Records
binary_choice	313,164
multi_choice	407,253
open	770,059

Attribute Coverage

This release contains 27 distinct attribute_type values, including:

• abnormality
• seizure_presence
• state
• eyes
• response_hand
• trial_type
• task_relevance
• task_variant
• tdcs_session
• transition
• word
• weight
• friction
• frequency
• emotion
• diagnosis
• several compositional attributes such as state+response_hand and trial_type+task_relevance

Loading the Dataset

import json

def load_jsonl(path):
    with open(path, "r", encoding="utf-8") as f:
        for line in f:
            yield json.loads(line)

for sample in load_jsonl("train.jsonl"):
    print(sample["task"], sample["question_type"], sample["question"])
    break

Representative Samples by Task, Question Type, and Task Family

The table below lists one representative example for every unique combination of task, question_type, and task_family. For choice-based questions, the example answer is expanded to the selected option text for readability.

Task	Question type	Task family	Example question	Example answer
asym	binary_choice	Contextual Semantic Decoding	What hand did the subject utilize to respond in this segment?	b. left hand
asym	multi_choice	Basic physiological state observation	What is the condition under which this segment was recorded?	a. emotion task
asym	multi_choice	High-level state inference	Which facial emotion is depicted in this segment?	b. happy (mouth open)
asym	multi_choice	Multi-step semantic reasoning	What characterizes this EEG segment in terms of both the experimental task and the hand used for the subject's response?	c. This segment was recorded during an emotion task in which the subject responded with the left hand.
asym	open	Basic physiological state observation	What is the condition under which this segment was recorded?	The segment was recorded during an emotion task.
asym	open	Contextual Semantic Decoding	Which hand did the subject use to respond during this segment?	The subject used the left hand to respond in this segment.
asym	open	High-level state inference	What facial emotion is expressed in this segment?	The facial emotion depicted is happy with an open mouth.
asym	open	Multi-step semantic reasoning	What characterizes this EEG segment in terms of both the experimental task and the hand used for the subject's response?	This segment was recorded during an emotion task in which the subject responded with the left hand.
bciciv	multi_choice	Motor intention and control command decoding	Which motor imagery class is being carried out in this segment?	a. left hand imagery
bciciv	open	Motor intention and control command decoding	Which motor imagery class is being carried out in this segment?	Left hand imagery is being carried out in this segment.
chbmit	binary_choice	Waveform and abnormal pattern detection	What is the status of the waveforms in this EEG segment?	b. seizure event
chbmit	open	Waveform and abnormal pattern detection	What is the status of the waveforms in this EEG segment?	The waveforms in this EEG segment represent a seizure event.
dementia	multi_choice	High-level state inference	To which clinical group does this resting EEG segment most likely belong?	a. Alzheimer's disease
dementia	open	High-level state inference	To which clinical group does this resting EEG segment most likely belong?	This resting EEG segment most likely belongs to a patient with Alzheimer's disease.
exo	binary_choice	Motor intention and control command decoding	What gait transition is taking place in this segment?	a. stop-to-walk transition
exo	multi_choice	Motor intention and control command decoding	Which gait transition is happening in this segment?	stop-to-walk transition
exo	open	Motor intention and control command decoding	What gait transition is taking place in this segment?	The gait transition taking place in this segment is a stop-to-walk transition.
gal	binary_choice	Basic physiological state observation	Does this segment originate from a resting state or a grasp-and-lift task?	b. grasp-and-lift task
gal	multi_choice	Contextual Semantic Decoding	What weight category is the object being lifted in this trial?	b. medium object (330g)
gal	multi_choice	Multi-step semantic reasoning	In terms of the behavioral state, object weight, and surface friction, what characterizes this EEG segment?	b. This EEG segment was recorded during a grasp-and-lift task involving a light object (165g) handled on a low friction surface (silk).
gal	open	Basic physiological state observation	What type of task does this EEG segment correspond to?	This EEG segment corresponds to a grasp-and-lift task.
gal	open	Contextual Semantic Decoding	What weight category is the object being lifted in this trial?	The object being lifted in this trial is a medium object weighing 330 grams.
gal	open	Multi-step semantic reasoning	In terms of the behavioral state, object weight, and surface friction, what characterizes this EEG segment?	This EEG segment was recorded during a grasp-and-lift task involving a light object (165g) handled on a low friction surface (silk).
is	binary_choice	Basic physiological state observation	Does this segment originate from a resting state or an inner speech task?	a. inner speech task
is	multi_choice	Contextual Semantic Decoding	In this EEG segment, which imagined word is represented?	e. ten
is	open	Basic physiological state observation	Does this segment originate from a resting state or an inner speech task?	This segment originates from an inner speech task.
is	open	Contextual Semantic Decoding	In this EEG segment, which imagined word is represented?	The imagined word represented in this EEG segment is 'ten'.
mom	binary_choice	Basic physiological state observation	Does this segment originate from a resting state or a task?	a. resting state
mom	binary_choice	Contextual Semantic Decoding	Is this trial classified as task-relevant or task-irrelevant?	a. task-relevant
mom	multi_choice	Contextual Semantic Decoding	What type of trial is this segment?	b. target
mom	multi_choice	Multi-step semantic reasoning	Given this EEG, what is the nature of the stimulus and its relevance to the subject's task?	b. This segment corresponds to a target stimulus that is task-relevant.
mom	open	Basic physiological state observation	Does this segment originate from a resting state or a task?	This segment originates from a resting state.
mom	open	Contextual Semantic Decoding	What type of trial does this EEG segment correspond to?	This segment corresponds to a target trial.
mom	open	Multi-step semantic reasoning	Given this EEG, what is the nature of the stimulus and its relevance to the subject's task?	This segment corresponds to a target stimulus that is task-relevant.
photo	binary_choice	Basic physiological state observation	Is this segment taken from a resting state or during photic stimulation?	b. resting state
photo	multi_choice	Contextual Semantic Decoding	What is the photic stimulation frequency applied in this segment?	a. 10 Hz
photo	open	Basic physiological state observation	During this photic resting-state segment, were the subject's eyes open or closed?	The subject's eyes were open.
photo	open	Contextual Semantic Decoding	What is the photic stimulation frequency applied in this segment?	The photic stimulation frequency applied in this segment is 10 Hz.
seed	multi_choice	High-level state inference	Which emotional condition best characterizes this EEG sample?	b. Neutral
seed	open	High-level state inference	What is the emotional state reflected in this EEG sample?	The emotional state reflected in this EEG sample is neutral.
surprise	binary_choice	Basic physiological state observation	Does this segment originate from a resting state or a task?	b. resting state
surprise	binary_choice	Contextual Semantic Decoding	What task variant is being executed in this segment?	b. SiB task
surprise	multi_choice	Contextual Semantic Decoding	What is the trial type within this segment?	a. dummy surprise
surprise	multi_choice	Multi-step semantic reasoning	Given this EEG, what is the combination of trial type and experimental task variant being performed?	a. This segment captures a dummy surprise trial carried out in the context of the SiB task variant.
surprise	open	Basic physiological state observation	What is the origin of this EEG segment-resting state or a specific task?	This EEG segment originates from a surprise task.
surprise	open	Contextual Semantic Decoding	Which variant of the task is being performed in this segment?	The SiB task is being performed in this segment.
surprise	open	Multi-step semantic reasoning	Given this EEG, what is the combination of trial type and experimental task variant being performed?	This segment captures a dummy surprise trial carried out in the context of the SiB task variant.
tdcs	binary_choice	Basic physiological state observation	Is this segment obtained during a resting state or a motor imagery task?	b. resting state
tdcs	multi_choice	Contextual Semantic Decoding	Which tDCS session corresponds to this segment?	a. anodal stimulation
tdcs	multi_choice	Multi-step semantic reasoning	Identify the experimental condition, specific motor imagery type, and stimulation session characterizing this EEG segment.	b. This EEG segment corresponds to a motor imagery task involving V1 imagery, performed during anodal stimulation.
tdcs	open	Basic physiological state observation	Was this EEG segment recorded while the subject was at rest or while performing a motor imagery task?	This EEG segment was recorded while the subject was at rest.
tdcs	open	Contextual Semantic Decoding	Which type of tDCS session does this EEG segment correspond to?	This segment corresponds to anodal stimulation.
tdcs	open	Multi-step semantic reasoning	What experimental condition, specific motor imagery type, and stimulation session characterize this EEG segment?	This EEG segment corresponds to a motor imagery task involving V1 imagery, performed during anodal stimulation.
tusz	binary_choice	Waveform and abnormal pattern detection	Is there electrographic evidence of seizure in this EEG clip?	b. No
tusz	multi_choice	Waveform and abnormal pattern detection	Which class best identifies this EEG interval as seizure or non-seizure?	c. Complex Partial Seizure
tusz	open	Waveform and abnormal pattern detection	What is the presence of electrographic seizure activity in this EEG clip?	There is no electrographic evidence of seizure in this EEG clip.
vrnf	binary_choice	Basic physiological state observation	Does this segment originate from a resting state or a neurofeedback task?	a. resting state
vrnf	open	Basic physiological state observation	What is the origin of this EEG segment-does it come from a resting state or a neurofeedback task?	This segment originates from a resting state.

Notes

• This README is derived from the released train.jsonl and test.jsonl files and from the benchmark design summary in EQA_OVERVIEW.md.
• The representative examples are intended to document coverage, not to define a train/test protocol beyond the provided split files.
• The file field points to the EEG segment source associated with each QA record.